From febc592af73db77c458726b755d85f6d1a9b5196 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=BB=84=E4=BC=9F=E9=9F=AC?= <huangweitao@fuzhi.ai>
Date: Fri, 16 Aug 2024 19:47:40 +0800
Subject: [PATCH 1/7] Force reply to human when finishing the current task

---
 metagpt/prompts/di/role_zero.py | 19 ++++++++++++
 metagpt/roles/di/role_zero.py   | 54 ++++++++++++++++++++++++++++-----
 metagpt/roles/di/swe_agent.py   | 35 ++++++++++-----------
 3 files changed, 83 insertions(+), 25 deletions(-)

diff --git a/metagpt/prompts/di/role_zero.py b/metagpt/prompts/di/role_zero.py
index 3c47ebf5f..f85128b38 100644
--- a/metagpt/prompts/di/role_zero.py
+++ b/metagpt/prompts/di/role_zero.py
@@ -218,3 +218,22 @@ QUICK_RESPONSE_SYSTEM_PROMPT = """
 {role_info}
 However, you MUST respond to the user message by yourself directly, DON'T ask your team members.
 """
+
+REPORT_TO_HUMAN_PROMPT = """
+# Current Plan
+{plan_status}
+
+Your have just finish a task, Use "RoleZero.reply_to_human" to report what you have done.
+The output format is :
+```json
+[
+    {{
+        "command_name": "RoleZero.reply_to_human",
+        "args": {{
+            "content": ""
+        }}
+    }}
+]
+```
+
+"""
diff --git a/metagpt/roles/di/role_zero.py b/metagpt/roles/di/role_zero.py
index c6cbfdfa8..727360ae9 100644
--- a/metagpt/roles/di/role_zero.py
+++ b/metagpt/roles/di/role_zero.py
@@ -26,6 +26,7 @@ from metagpt.prompts.di.role_zero import (
     QUICK_THINK_PROMPT,
     QUICK_THINK_SYSTEM_PROMPT,
     REGENERATE_PROMPT,
+    REPORT_TO_HUMAN_PROMPT,
     ROLE_INSTRUCTION,
     SYSTEM_PROMPT,
     THOUGHT_GUIDANCE,
@@ -86,6 +87,8 @@ class RoleZero(Role):
     use_fixed_sop: bool = False
     requirements_constraints: str = ""  # the constraints in user requirements
 
+    command_history: list[str] = []
+
     @model_validator(mode="after")
     def set_plan_and_tool(self) -> "RoleZero":
         # We force using this parameter for DataAnalyst
@@ -234,21 +237,57 @@ class RoleZero(Role):
         if self.use_fixed_sop:
             return await super()._act()
 
-        commands, ok = await self._parse_commands()
+        commands, ok = await self._parse_commands(self.command_rsp)
         if not ok:
             error_msg = commands
+            self.rc.memory.add(UserMessage(content=error_msg))
             return error_msg
         logger.info(f"Commands: \n{commands}")
         outputs = await self._run_commands(commands)
         logger.info(f"Commands outputs: \n{outputs}")
         self.rc.memory.add(UserMessage(content=outputs))
 
+        # Report what is done when finishing the task.
+        current_command_list = [command["command_name"] for command in commands]
+        self.command_history.extend(current_command_list)
+        if self.check_whether_report_to_human():
+            memory = self.rc.memory.get(self.memory_k)
+            memory = await self.parse_browser_actions(memory)
+            memory = self.parse_images(memory)
+            plan_status, _ = self._get_plan_status()
+            prompt = REPORT_TO_HUMAN_PROMPT.format(plan_status=plan_status)
+            req = self.llm.format_msg(memory + [UserMessage(content=prompt)])
+            respond_command = await self.llm.aask(msg=req)
+            commands, ok = await self._parse_commands(respond_command)
+            if ok and len(commands) == 1 and commands[0]["command_name"] == "RoleZero.reply_to_human":
+                cmd = commands[0]
+                report_result = await self.reply_to_human(cmd["args"])
+                self.rc.memory.add(AIMessage(content=respond_command))
+                self.rc.memory.add(UserMessage(content=report_result))
+                self.command_history.append("RoleZero.reply_to_human")
+                logger.info(f"Commands outputs: \n{report_result}")
+                outputs += report_result
+
         return AIMessage(
             content=f"I have finished the task, please mark my task as finished. Outputs: {outputs}",
             sent_from=self.name,
             cause_by=RunCommand,
         )
 
+    def check_whether_report_to_human(self):
+        """ "Check whether add reply to human command when finish current task"""
+        interaction_with_human = ["RoleZero.ask_human", "RoleZero.reply_to_human"]
+        plan_end_action = ["Plan.finish_current_task", "end"]
+        flag = 0
+        for command_name in self.command_history[::-1]:
+            if command_name in plan_end_action:
+                flag |= 1
+            elif command_name in interaction_with_human:
+                flag |= 2
+            else:
+                break
+        return flag == 1
+
     async def _react(self) -> Message:
         # NOTE: Diff 1: Each time landing here means news is observed, set todo to allow news processing in _think
         self._set_state(0)
@@ -332,7 +371,7 @@ class RoleZero(Role):
             command_rsp = await self.llm.aask(regenerate_req)
         return command_rsp
 
-    async def _parse_commands(self) -> Tuple[List[Dict], bool]:
+    async def _parse_commands(self, command_rsp) -> Tuple[List[Dict], bool]:
         """Retrieves commands from the Large Language Model (LLM).
 
         This function attempts to retrieve a list of commands from the LLM by
@@ -344,20 +383,20 @@ class RoleZero(Role):
                 - A boolean flag indicating success (True) or failure (False).
         """
         try:
-            commands = CodeParser.parse_code(block=None, lang="json", text=self.command_rsp)
+            commands = CodeParser.parse_code(block=None, lang="json", text=command_rsp)
             if commands.endswith("]") and not commands.startswith("["):
                 commands = "[" + commands
             commands = json.loads(repair_llm_raw_output(output=commands, req_keys=[None], repair_type=RepairType.JSON))
         except json.JSONDecodeError as e:
-            logger.warning(f"Failed to parse JSON for: {self.command_rsp}. Trying to repair...")
+            logger.warning(f"Failed to parse JSON for: {command_rsp}. Trying to repair...")
             commands = await self.llm.aask(
-                msg=JSON_REPAIR_PROMPT.format(json_data=self.command_rsp, json_decode_error=str(e))
+                msg=JSON_REPAIR_PROMPT.format(json_data=command_rsp, json_decode_error=str(e))
             )
             try:
                 commands = json.loads(CodeParser.parse_code(block=None, lang="json", text=commands))
             except json.JSONDecodeError:
                 # repair escape error of code and math
-                commands = CodeParser.parse_code(block=None, lang="json", text=self.command_rsp)
+                commands = CodeParser.parse_code(block=None, lang="json", text=command_rsp)
                 new_command = repair_escape_error(commands)
                 commands = json.loads(
                     repair_llm_raw_output(output=new_command, req_keys=[None], repair_type=RepairType.JSON)
@@ -365,8 +404,7 @@ class RoleZero(Role):
         except Exception as e:
             tb = traceback.format_exc()
             print(tb)
-            error_msg = UserMessage(content=str(e))
-            self.rc.memory.add(error_msg)
+            error_msg = str(e)
             return error_msg, False
 
         # 为了对LLM不按格式生成进行容错
diff --git a/metagpt/roles/di/swe_agent.py b/metagpt/roles/di/swe_agent.py
index e1d2c9613..e90fc3045 100644
--- a/metagpt/roles/di/swe_agent.py
+++ b/metagpt/roles/di/swe_agent.py
@@ -9,6 +9,7 @@ from metagpt.prompts.di.swe_agent import (
     NEXT_STEP_TEMPLATE,
 )
 from metagpt.roles.di.role_zero import RoleZero
+from metagpt.schema import Message
 from metagpt.tools.libs.git import git_create_pull
 from metagpt.tools.libs.terminal import Bash
 
@@ -32,8 +33,6 @@ class SWEAgent(RoleZero):
     async def _think(self) -> bool:
         await self._format_instruction()
         res = await super()._think()
-        if self.run_eval:
-            await self._parse_commands_for_eval()
         return res
 
     def _update_tool_execution(self):
@@ -55,6 +54,12 @@ class SWEAgent(RoleZero):
         bash_state = json.loads(state_output)
         self.cmd_prompt_current_state = CURRENT_BASH_STATE.format(**bash_state).strip()
 
+    async def _act(self) -> Message:
+        message = await super()._act()
+        if self.run_eval:
+            self._parse_commands_for_eval()
+        return message
+
     async def _parse_commands_for_eval(self):
         """
         Handles actions based on parsed commands.
@@ -65,23 +70,19 @@ class SWEAgent(RoleZero):
         This function is specifically added for SWE bench evaluation.
         """
         # only import when evaluation is needed
-        from metagpt.tools.swe_agent_commands.swe_agent_utils import extract_patch
+        if not self.rc.todo:
+            from metagpt.tools.swe_agent_commands.swe_agent_utils import extract_patch
 
-        commands, ok = await self._parse_commands()
-        if not ok:
-            return
-        for cmd in commands:
-            if "end" != cmd.get("command_name", ""):
-                return
-        try:
-            diff_output = await self.terminal.run("git diff --cached")
-            clear_diff = extract_patch(diff_output)
-            logger.info(f"Diff output: \n{clear_diff}")
-            if clear_diff:
-                self.output_diff = clear_diff
+            # swe agent have been stop. it means  'end' or other command which can stop the swe agent have been executed
+            try:
+                diff_output = await self.terminal.run("git diff --cached")
+                clear_diff = extract_patch(diff_output)
+                logger.info(f"Diff output: \n{clear_diff}")
+                if clear_diff:
+                    self.output_diff = clear_diff
 
-        except Exception as e:
-            logger.error(f"Error during submission: {e}")
+            except Exception as e:
+                logger.error(f"Error during submission: {e}")
 
     def _retrieve_experience(self) -> str:
         return MINIMAL_EXAMPLE

From 0246b659b7f7448ef0545697742e89066b44b1d5 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=BB=84=E4=BC=9F=E9=9F=AC?= <huangweitao@fuzhi.ai>
Date: Tue, 20 Aug 2024 17:13:15 +0800
Subject: [PATCH 2/7] Add a summary after the agent stops

---
 metagpt/prompts/di/role_zero.py | 27 ++++++++--------
 metagpt/roles/di/role_zero.py   | 57 ++++++++++++---------------------
 metagpt/roles/di/swe_agent.py   |  3 +-
 metagpt/roles/di/team_leader.py |  2 ++
 4 files changed, 37 insertions(+), 52 deletions(-)

diff --git a/metagpt/prompts/di/role_zero.py b/metagpt/prompts/di/role_zero.py
index f85128b38..2294fa731 100644
--- a/metagpt/prompts/di/role_zero.py
+++ b/metagpt/prompts/di/role_zero.py
@@ -220,20 +220,19 @@ However, you MUST respond to the user message by yourself directly, DON'T ask yo
 """
 
 REPORT_TO_HUMAN_PROMPT = """
-# Current Plan
-{plan_status}
+# Restrictions
+{requirements_constraints}
 
-Your have just finish a task, Use "RoleZero.reply_to_human" to report what you have done.
-The output format is :
-```json
-[
-    {{
-        "command_name": "RoleZero.reply_to_human",
-        "args": {{
-            "content": ""
-        }}
-    }}
-]
-```
+Your have just finish all tasks, Use "RoleZero.reply_to_human" to answer the user requirements.
+Report to human what you have done. Do Not ouput any other format.
+Your reply is:
 
 """
+SUMMARY_PROMPY = """
+# Restrictions
+{requirements_constraints}
+
+You have just completed some tasks. 
+Summarize the tasks you have accomplished without including detailed information. 
+If there are any deliverables, list their descriptions and provide their file paths.
+"""
diff --git a/metagpt/roles/di/role_zero.py b/metagpt/roles/di/role_zero.py
index 727360ae9..48967d7ef 100644
--- a/metagpt/roles/di/role_zero.py
+++ b/metagpt/roles/di/role_zero.py
@@ -28,6 +28,7 @@ from metagpt.prompts.di.role_zero import (
     REGENERATE_PROMPT,
     REPORT_TO_HUMAN_PROMPT,
     ROLE_INSTRUCTION,
+    SUMMARY_PROMPY,
     SYSTEM_PROMPT,
     THOUGHT_GUIDANCE,
 )
@@ -68,6 +69,8 @@ class RoleZero(Role):
     react_mode: Literal["react"] = "react"
     max_react_loop: int = 20  # used for react mode
 
+    # Summary Mode
+    use_summary: bool = True
     # Tools
     tools: list[str] = []  # Use special symbol ["<all>"] to indicate use of all registered tools
     tool_recommender: Optional[ToolRecommender] = None
@@ -87,8 +90,6 @@ class RoleZero(Role):
     use_fixed_sop: bool = False
     requirements_constraints: str = ""  # the constraints in user requirements
 
-    command_history: list[str] = []
-
     @model_validator(mode="after")
     def set_plan_and_tool(self) -> "RoleZero":
         # We force using this parameter for DataAnalyst
@@ -247,26 +248,24 @@ class RoleZero(Role):
         logger.info(f"Commands outputs: \n{outputs}")
         self.rc.memory.add(UserMessage(content=outputs))
 
-        # Report what is done when finishing the task.
-        current_command_list = [command["command_name"] for command in commands]
-        self.command_history.extend(current_command_list)
-        if self.check_whether_report_to_human():
-            memory = self.rc.memory.get(self.memory_k)
-            memory = await self.parse_browser_actions(memory)
-            memory = self.parse_images(memory)
-            plan_status, _ = self._get_plan_status()
-            prompt = REPORT_TO_HUMAN_PROMPT.format(plan_status=plan_status)
-            req = self.llm.format_msg(memory + [UserMessage(content=prompt)])
-            respond_command = await self.llm.aask(msg=req)
-            commands, ok = await self._parse_commands(respond_command)
-            if ok and len(commands) == 1 and commands[0]["command_name"] == "RoleZero.reply_to_human":
-                cmd = commands[0]
-                report_result = await self.reply_to_human(cmd["args"])
-                self.rc.memory.add(AIMessage(content=respond_command))
-                self.rc.memory.add(UserMessage(content=report_result))
-                self.command_history.append("RoleZero.reply_to_human")
-                logger.info(f"Commands outputs: \n{report_result}")
-                outputs += report_result
+        if any(["end" == command["command_name"] for command in commands]):
+            # Ensure reply to the human before the "end" command is executed.
+            if all(["reply_to_human" not in memory.content for memory in self.get_memories(k=5)]):
+                memory = self.rc.memory.get(self.memory_k)
+                reply_to_human_prompt = REPORT_TO_HUMAN_PROMPT.format(
+                    requirements_constraints=self.requirements_constraints,
+                )
+                reply_content = await self.llm.aask(self.llm.format_msg(memory + [UserMessage(reply_to_human_prompt)]))
+                await self.reply_to_human(content=reply_content)
+                self.rc.memory.add(AIMessage(content=reply_content, cause_by=RunCommand))
+
+            # Summary of the Completed Task and Deliverables
+            if self.use_summary:
+                memory = self.rc.memory.get(self.memory_k)
+                summary_prompt = SUMMARY_PROMPY.format(
+                    requirements_constraints=self.requirements_constraints,
+                )
+                outputs = await self.llm.aask(self.llm.format_msg(memory + [UserMessage(summary_prompt)]))
 
         return AIMessage(
             content=f"I have finished the task, please mark my task as finished. Outputs: {outputs}",
@@ -274,20 +273,6 @@ class RoleZero(Role):
             cause_by=RunCommand,
         )
 
-    def check_whether_report_to_human(self):
-        """ "Check whether add reply to human command when finish current task"""
-        interaction_with_human = ["RoleZero.ask_human", "RoleZero.reply_to_human"]
-        plan_end_action = ["Plan.finish_current_task", "end"]
-        flag = 0
-        for command_name in self.command_history[::-1]:
-            if command_name in plan_end_action:
-                flag |= 1
-            elif command_name in interaction_with_human:
-                flag |= 2
-            else:
-                break
-        return flag == 1
-
     async def _react(self) -> Message:
         # NOTE: Diff 1: Each time landing here means news is observed, set todo to allow news processing in _think
         self._set_state(0)
diff --git a/metagpt/roles/di/swe_agent.py b/metagpt/roles/di/swe_agent.py
index e90fc3045..d54c9dd44 100644
--- a/metagpt/roles/di/swe_agent.py
+++ b/metagpt/roles/di/swe_agent.py
@@ -69,11 +69,10 @@ class SWEAgent(RoleZero):
 
         This function is specifically added for SWE bench evaluation.
         """
-        # only import when evaluation is needed
+        # If todo switches to None, it indicates that this is the final round of reactions, and the Swe-Agent will stop. Use git diff to store any changes made.
         if not self.rc.todo:
             from metagpt.tools.swe_agent_commands.swe_agent_utils import extract_patch
 
-            # swe agent have been stop. it means  'end' or other command which can stop the swe agent have been executed
             try:
                 diff_output = await self.terminal.run("git diff --cached")
                 clear_diff = extract_patch(diff_output)
diff --git a/metagpt/roles/di/team_leader.py b/metagpt/roles/di/team_leader.py
index 97100f295..112ca5a84 100644
--- a/metagpt/roles/di/team_leader.py
+++ b/metagpt/roles/di/team_leader.py
@@ -30,6 +30,8 @@ class TeamLeader(RoleZero):
 
     experience_retriever: Annotated[ExpRetriever, Field(exclude=True)] = SimpleExpRetriever()
 
+    use_summary: bool = False
+
     def _update_tool_execution(self):
         self.tool_execution_map.update(
             {

From 502a0e2ad6e784af6668d4d2a119b6152a695128 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=BB=84=E4=BC=9F=E9=9F=AC?= <huangweitao@fuzhi.ai>
Date: Tue, 20 Aug 2024 19:21:17 +0800
Subject: [PATCH 3/7] combine summary and last reply to human

---
 metagpt/prompts/di/role_zero.py |  6 +++--
 metagpt/roles/di/role_zero.py   | 48 ++++++++++++++++++---------------
 2 files changed, 30 insertions(+), 24 deletions(-)

diff --git a/metagpt/prompts/di/role_zero.py b/metagpt/prompts/di/role_zero.py
index 2294fa731..80a505d17 100644
--- a/metagpt/prompts/di/role_zero.py
+++ b/metagpt/prompts/di/role_zero.py
@@ -233,6 +233,8 @@ SUMMARY_PROMPY = """
 {requirements_constraints}
 
 You have just completed some tasks. 
-Summarize the tasks you have accomplished without including detailed information. 
-If there are any deliverables, list their descriptions and provide their file paths.
+You must reply to human what you have done and what is the result.
+If there are any deliverables, provide descriptions and file paths.
+Do Not ouput any other format. limited in 300 words.
+Your reply is:
 """
diff --git a/metagpt/roles/di/role_zero.py b/metagpt/roles/di/role_zero.py
index 48967d7ef..4ab70409e 100644
--- a/metagpt/roles/di/role_zero.py
+++ b/metagpt/roles/di/role_zero.py
@@ -26,7 +26,6 @@ from metagpt.prompts.di.role_zero import (
     QUICK_THINK_PROMPT,
     QUICK_THINK_SYSTEM_PROMPT,
     REGENERATE_PROMPT,
-    REPORT_TO_HUMAN_PROMPT,
     ROLE_INSTRUCTION,
     SUMMARY_PROMPY,
     SYSTEM_PROMPT,
@@ -248,25 +247,6 @@ class RoleZero(Role):
         logger.info(f"Commands outputs: \n{outputs}")
         self.rc.memory.add(UserMessage(content=outputs))
 
-        if any(["end" == command["command_name"] for command in commands]):
-            # Ensure reply to the human before the "end" command is executed.
-            if all(["reply_to_human" not in memory.content for memory in self.get_memories(k=5)]):
-                memory = self.rc.memory.get(self.memory_k)
-                reply_to_human_prompt = REPORT_TO_HUMAN_PROMPT.format(
-                    requirements_constraints=self.requirements_constraints,
-                )
-                reply_content = await self.llm.aask(self.llm.format_msg(memory + [UserMessage(reply_to_human_prompt)]))
-                await self.reply_to_human(content=reply_content)
-                self.rc.memory.add(AIMessage(content=reply_content, cause_by=RunCommand))
-
-            # Summary of the Completed Task and Deliverables
-            if self.use_summary:
-                memory = self.rc.memory.get(self.memory_k)
-                summary_prompt = SUMMARY_PROMPY.format(
-                    requirements_constraints=self.requirements_constraints,
-                )
-                outputs = await self.llm.aask(self.llm.format_msg(memory + [UserMessage(summary_prompt)]))
-
         return AIMessage(
             content=f"I have finished the task, please mark my task as finished. Outputs: {outputs}",
             sent_from=self.name,
@@ -442,8 +422,7 @@ class RoleZero(Role):
             command_output = "Current task is finished. If all tasks are finished, use 'end' to stop."
 
         elif cmd["command_name"] == "end":
-            self._set_state(-1)
-            command_output = ""
+            command_output = await self._end()
 
         # output from bash.run may be empty, add decorations to the output to ensure visibility.
         elif cmd["command_name"] == "Bash.run":
@@ -502,3 +481,28 @@ class RoleZero(Role):
         if not isinstance(self.rc.env, MGXEnv):
             return "Not in MGXEnv, command will not be executed."
         return await self.rc.env.reply_to_human(content, sent_from=self)
+
+    async def _end(self):
+        # summary
+        memory = self.get_memories(k=self.memory_k)
+        summary_prompt = SUMMARY_PROMPY.format(
+            requirements_constraints=self.requirements_constraints,
+        )
+        reply_content = await self.llm.aask(self.llm.format_msg(memory + [UserMessage(summary_prompt)]))
+        # Ensure reply to the human before the "end" command is executed.
+        need_reply = True
+        for memory in self.get_memories(k=5)[::-1]:
+            if "reply_to_human" in memory.content:
+                need_reply = False
+                break
+            if "[Message]" in memory.content:
+                # Receive new message from other
+                need_reply = True
+                break
+        if need_reply:
+            await self.reply_to_human(content=reply_content)
+            self.rc.memory.add(AIMessage(content=reply_content, cause_by=RunCommand))
+
+        self._set_state(-1)
+
+        return f"{reply_content} \n Plan.end executed Task is finished"

From 1730897730962dcaeb438a62ca14da2decaf7d82 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=BB=84=E4=BC=9F=E9=9F=AC?= <huangweitao@fuzhi.ai>
Date: Tue, 20 Aug 2024 19:45:13 +0800
Subject: [PATCH 4/7] fix issues

---
 metagpt/roles/di/role_zero.py   | 8 +++-----
 metagpt/roles/di/team_leader.py | 2 --
 2 files changed, 3 insertions(+), 7 deletions(-)

diff --git a/metagpt/roles/di/role_zero.py b/metagpt/roles/di/role_zero.py
index 4ab70409e..59714b726 100644
--- a/metagpt/roles/di/role_zero.py
+++ b/metagpt/roles/di/role_zero.py
@@ -68,8 +68,6 @@ class RoleZero(Role):
     react_mode: Literal["react"] = "react"
     max_react_loop: int = 20  # used for react mode
 
-    # Summary Mode
-    use_summary: bool = True
     # Tools
     tools: list[str] = []  # Use special symbol ["<all>"] to indicate use of all registered tools
     tool_recommender: Optional[ToolRecommender] = None
@@ -483,6 +481,8 @@ class RoleZero(Role):
         return await self.rc.env.reply_to_human(content, sent_from=self)
 
     async def _end(self):
+        self._set_state(-1)
+
         # summary
         memory = self.get_memories(k=self.memory_k)
         summary_prompt = SUMMARY_PROMPY.format(
@@ -503,6 +503,4 @@ class RoleZero(Role):
             await self.reply_to_human(content=reply_content)
             self.rc.memory.add(AIMessage(content=reply_content, cause_by=RunCommand))
 
-        self._set_state(-1)
-
-        return f"{reply_content} \n Plan.end executed Task is finished"
+        return reply_content
diff --git a/metagpt/roles/di/team_leader.py b/metagpt/roles/di/team_leader.py
index 112ca5a84..97100f295 100644
--- a/metagpt/roles/di/team_leader.py
+++ b/metagpt/roles/di/team_leader.py
@@ -30,8 +30,6 @@ class TeamLeader(RoleZero):
 
     experience_retriever: Annotated[ExpRetriever, Field(exclude=True)] = SimpleExpRetriever()
 
-    use_summary: bool = False
-
     def _update_tool_execution(self):
         self.tool_execution_map.update(
             {

From beb66b05b42ca8731f341edc652f84fa6d0df352 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=BB=84=E4=BC=9F=E9=9F=AC?= <huangweitao@fuzhi.ai>
Date: Tue, 20 Aug 2024 20:41:47 +0800
Subject: [PATCH 5/7] decompose reply to human and summary

---
 metagpt/prompts/di/role_zero.py | 11 +++++-----
 metagpt/roles/di/role_zero.py   | 37 ++++++++++++++++-----------------
 metagpt/roles/di/team_leader.py |  2 ++
 3 files changed, 25 insertions(+), 25 deletions(-)

diff --git a/metagpt/prompts/di/role_zero.py b/metagpt/prompts/di/role_zero.py
index 80a505d17..caba52e57 100644
--- a/metagpt/prompts/di/role_zero.py
+++ b/metagpt/prompts/di/role_zero.py
@@ -223,8 +223,9 @@ REPORT_TO_HUMAN_PROMPT = """
 # Restrictions
 {requirements_constraints}
 
-Your have just finish all tasks, Use "RoleZero.reply_to_human" to answer the user requirements.
-Report to human what you have done. Do Not ouput any other format.
+You have just finished all tasks.
+Reply to the human requirements.
+Do not output any other format.
 Your reply is:
 
 """
@@ -233,8 +234,6 @@ SUMMARY_PROMPY = """
 {requirements_constraints}
 
 You have just completed some tasks. 
-You must reply to human what you have done and what is the result.
-If there are any deliverables, provide descriptions and file paths.
-Do Not ouput any other format. limited in 300 words.
-Your reply is:
+Summarize the tasks you have accomplished without including detailed information. 
+If there are any deliverables, list their descriptions and provide their file paths.
 """
diff --git a/metagpt/roles/di/role_zero.py b/metagpt/roles/di/role_zero.py
index 59714b726..8dfd888d2 100644
--- a/metagpt/roles/di/role_zero.py
+++ b/metagpt/roles/di/role_zero.py
@@ -26,6 +26,7 @@ from metagpt.prompts.di.role_zero import (
     QUICK_THINK_PROMPT,
     QUICK_THINK_SYSTEM_PROMPT,
     REGENERATE_PROMPT,
+    REPORT_TO_HUMAN_PROMPT,
     ROLE_INSTRUCTION,
     SUMMARY_PROMPY,
     SYSTEM_PROMPT,
@@ -68,6 +69,8 @@ class RoleZero(Role):
     react_mode: Literal["react"] = "react"
     max_react_loop: int = 20  # used for react mode
 
+    # Summary Mode
+    use_summary: bool = True
     # Tools
     tools: list[str] = []  # Use special symbol ["<all>"] to indicate use of all registered tools
     tool_recommender: Optional[ToolRecommender] = None
@@ -482,25 +485,21 @@ class RoleZero(Role):
 
     async def _end(self):
         self._set_state(-1)
-
-        # summary
-        memory = self.get_memories(k=self.memory_k)
-        summary_prompt = SUMMARY_PROMPY.format(
-            requirements_constraints=self.requirements_constraints,
-        )
-        reply_content = await self.llm.aask(self.llm.format_msg(memory + [UserMessage(summary_prompt)]))
         # Ensure reply to the human before the "end" command is executed.
-        need_reply = True
-        for memory in self.get_memories(k=5)[::-1]:
-            if "reply_to_human" in memory.content:
-                need_reply = False
-                break
-            if "[Message]" in memory.content:
-                # Receive new message from other
-                need_reply = True
-                break
-        if need_reply:
+        if not any(["reply_to_human" in memory.content for memory in self.get_memories(k=5)]):
+            memory = self.rc.memory.get(self.memory_k)
+            reply_to_human_prompt = REPORT_TO_HUMAN_PROMPT.format(
+                requirements_constraints=self.requirements_constraints,
+            )
+            reply_content = await self.llm.aask(self.llm.format_msg(memory + [UserMessage(reply_to_human_prompt)]))
             await self.reply_to_human(content=reply_content)
             self.rc.memory.add(AIMessage(content=reply_content, cause_by=RunCommand))
-
-        return reply_content
+        outputs = ""
+        # Summary of the Completed Task and Deliverables
+        if self.use_summary:
+            memory = self.rc.memory.get(self.memory_k)
+            summary_prompt = SUMMARY_PROMPY.format(
+                requirements_constraints=self.requirements_constraints,
+            )
+            outputs = await self.llm.aask(self.llm.format_msg(memory + [UserMessage(summary_prompt)]))
+        return outputs
diff --git a/metagpt/roles/di/team_leader.py b/metagpt/roles/di/team_leader.py
index 97100f295..112ca5a84 100644
--- a/metagpt/roles/di/team_leader.py
+++ b/metagpt/roles/di/team_leader.py
@@ -30,6 +30,8 @@ class TeamLeader(RoleZero):
 
     experience_retriever: Annotated[ExpRetriever, Field(exclude=True)] = SimpleExpRetriever()
 
+    use_summary: bool = False
+
     def _update_tool_execution(self):
         self.tool_execution_map.update(
             {

From ca66af381a8fedb1faadc3fe022dcd8a5819a877 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=BB=84=E4=BC=9F=E9=9F=AC?= <huangweitao@fuzhi.ai>
Date: Tue, 20 Aug 2024 20:52:08 +0800
Subject: [PATCH 6/7] fix format issues

---
 metagpt/prompts/di/role_zero.py | 2 +-
 metagpt/roles/di/role_zero.py   | 7 +++----
 2 files changed, 4 insertions(+), 5 deletions(-)

diff --git a/metagpt/prompts/di/role_zero.py b/metagpt/prompts/di/role_zero.py
index caba52e57..ce8c61dd5 100644
--- a/metagpt/prompts/di/role_zero.py
+++ b/metagpt/prompts/di/role_zero.py
@@ -229,7 +229,7 @@ Do not output any other format.
 Your reply is:
 
 """
-SUMMARY_PROMPY = """
+SUMMARY_PROMPT = """
 # Restrictions
 {requirements_constraints}
 
diff --git a/metagpt/roles/di/role_zero.py b/metagpt/roles/di/role_zero.py
index 8dfd888d2..4e61ceab3 100644
--- a/metagpt/roles/di/role_zero.py
+++ b/metagpt/roles/di/role_zero.py
@@ -28,7 +28,7 @@ from metagpt.prompts.di.role_zero import (
     REGENERATE_PROMPT,
     REPORT_TO_HUMAN_PROMPT,
     ROLE_INSTRUCTION,
-    SUMMARY_PROMPY,
+    SUMMARY_PROMPT,
     SYSTEM_PROMPT,
     THOUGHT_GUIDANCE,
 )
@@ -485,9 +485,9 @@ class RoleZero(Role):
 
     async def _end(self):
         self._set_state(-1)
+        memory = self.rc.memory.get(self.memory_k)
         # Ensure reply to the human before the "end" command is executed.
         if not any(["reply_to_human" in memory.content for memory in self.get_memories(k=5)]):
-            memory = self.rc.memory.get(self.memory_k)
             reply_to_human_prompt = REPORT_TO_HUMAN_PROMPT.format(
                 requirements_constraints=self.requirements_constraints,
             )
@@ -497,8 +497,7 @@ class RoleZero(Role):
         outputs = ""
         # Summary of the Completed Task and Deliverables
         if self.use_summary:
-            memory = self.rc.memory.get(self.memory_k)
-            summary_prompt = SUMMARY_PROMPY.format(
+            summary_prompt = SUMMARY_PROMPT.format(
                 requirements_constraints=self.requirements_constraints,
             )
             outputs = await self.llm.aask(self.llm.format_msg(memory + [UserMessage(summary_prompt)]))

From 14642407d15eee8e1522f5496856124038cf0f4a Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=BB=84=E4=BC=9F=E9=9F=AC?= <huangweitao@fuzhi.ai>
Date: Tue, 20 Aug 2024 20:55:33 +0800
Subject: [PATCH 7/7] fix format issues

---
 metagpt/roles/di/role_zero.py | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/metagpt/roles/di/role_zero.py b/metagpt/roles/di/role_zero.py
index 4e61ceab3..1153cb22d 100644
--- a/metagpt/roles/di/role_zero.py
+++ b/metagpt/roles/di/role_zero.py
@@ -69,8 +69,6 @@ class RoleZero(Role):
     react_mode: Literal["react"] = "react"
     max_react_loop: int = 20  # used for react mode
 
-    # Summary Mode
-    use_summary: bool = True
     # Tools
     tools: list[str] = []  # Use special symbol ["<all>"] to indicate use of all registered tools
     tool_recommender: Optional[ToolRecommender] = None
@@ -89,6 +87,7 @@ class RoleZero(Role):
     memory_k: int = 20  # number of memories (messages) to use as historical context
     use_fixed_sop: bool = False
     requirements_constraints: str = ""  # the constraints in user requirements
+    use_summary: bool = True  # whether to summarize at the end
 
     @model_validator(mode="after")
     def set_plan_and_tool(self) -> "RoleZero":