add quick think

metagpt/prompts/di/role_zero.py

@@ -51,6 +51,7 @@ Some text indicating your thoughts, such as how you should update the plan statu
 ```
 Notice: your output JSON data section must start with **```json [**
 """
+
 JSON_REPAIR_PROMPT = """
 ## json data
 {json_data}
@@ -61,3 +62,10 @@ Formatted JSON data
 ```
 Help check if there are any formatting issues with the JSON data? If so, please help format it
 """
+
+QUICK_THINK_PROMPT = """
+Decide if the latest user message is a quick question.
+Quick questions include common-sense, logical, math questions, greetings, or casual chat that you can answer directly, excluding software development tasks.
+Respond with "#YES#, (then start your actual response to the question...)" if so, otherwise, simply respond with "#NO#".
+Your response:
+"""

metagpt/roles/di/role_zero.py

@@ -8,12 +8,13 @@ from typing import Callable, Dict, List, Literal, Tuple
 
 from pydantic import model_validator
 
-from metagpt.actions import Action
+from metagpt.actions import Action, UserRequirement
 from metagpt.actions.di.run_command import RunCommand
 from metagpt.logs import logger
 from metagpt.prompts.di.role_zero import (
     CMD_PROMPT,
     JSON_REPAIR_PROMPT,
+    QUICK_THINK_PROMPT,
     ROLE_INSTRUCTION,
 )
 from metagpt.roles import Role
@@ -24,7 +25,7 @@ from metagpt.tools.libs.browser import Browser
 from metagpt.tools.libs.editor import Editor
 from metagpt.tools.tool_recommend import BM25ToolRecommender, ToolRecommender
 from metagpt.tools.tool_registry import register_tool
-from metagpt.utils.common import CodeParser
+from metagpt.utils.common import CodeParser, any_to_str
 from metagpt.utils.repair_llm_raw_output import RepairType, repair_llm_raw_output
 from metagpt.utils.report import ThoughtReporter
 
@@ -47,7 +48,7 @@ class RoleZero(Role):
 
     # Tools
     tools: list[str] = []  # Use special symbol ["<all>"] to indicate use of all registered tools
-    tool_recommender: ToolRecommender = None
+    tool_recommender: ToolRecommender = ToolRecommender()
     tool_execution_map: dict[str, Callable] = {}
     special_tool_commands: list[str] = ["Plan.finish_current_task", "end"]
     # Equipped with three basic tools by default for optional use
@@ -183,9 +184,14 @@ class RoleZero(Role):
         )
 
     async def _react(self) -> Message:
-        # NOTE: Diff 1: Each time landing here means observing news, set todo to allow news processing in _think
+        # NOTE: Diff 1: Each time landing here means news is observed, set todo to allow news processing in _think
         self._set_state(0)
 
+        # problems solvable by quick thinking doesn't need to a formal think-act cycle
+        quick_rsp = await self._quick_think()
+        if quick_rsp:
+            return quick_rsp
+
         actions_taken = 0
         rsp = AIMessage(content="No actions taken yet", cause_by=Action)  # will be overwritten after Role _act
         while actions_taken < self.rc.max_react_loop:
@@ -202,6 +208,29 @@ class RoleZero(Role):
             actions_taken += 1
         return rsp  # return output from the last action
 
+    async def _quick_think(self) -> Message:
+        msg = self.rc.news[-1]
+        rsp_msg = None
+        if msg.cause_by != any_to_str(UserRequirement):
+            # Agents themselves won't generate quick questions, use this rule to reduce extra llm calls
+            return rsp_msg
+
+        context = self.llm.format_msg(self.get_memories(k=4) + [UserMessage(content=QUICK_THINK_PROMPT)])
+        rsp = await self.llm.aask(context)
+
+        pattern = r"#YES#,? ?"
+        if re.search(pattern, rsp):
+            answer = re.sub(pattern, "", rsp).strip()
+            self.rc.memory.add(AIMessage(content=answer, cause_by=RunCommand))
+            await self.reply_to_human(content=answer)
+            rsp_msg = AIMessage(
+                content="Complete run",
+                sent_from=self.name,
+                cause_by=RunCommand,
+            )
+
+        return rsp_msg
+
     async def _parse_commands(self) -> Tuple[List[Dict], bool]:
         """Retrieves commands from the Large Language Model (LLM).
 

tests/metagpt/environment/mgx_env/run_mgx_env.py

@@ -129,9 +129,11 @@ TL_CHAT2 = """Solve the issue at this link"""  # expecting clarification
 TL_CHAT3 = """Who is the first man landing on Moon"""  # expecting answering directly
 TL_CHAT4 = """Find all zeros in the indicated finite field of the given polynomial with coefficients in that field. x^5 + 3x^3 + x^2 + 2x in Z_5"""  # expecting answering directly
 TL_CHAT5 = """Find the degree for the given field extension Q(sqrt(2), sqrt(3), sqrt(18)) over Q."""  # expecting answering directly
-TL_CHAT6 = """Statement 1 | A ring homomorphism is one to one if and only if the kernel is {{0}},. Statement 2 | Q is an ideal in R"""  # expecting answering directly
+TL_CHAT6 = """True or False? Statement 1 | A ring homomorphism is one to one if and only if the kernel is {{0}},. Statement 2 | Q is an ideal in R"""  # expecting answering directly
 TL_CHAT7 = """Jean has 30 lollipops. Jean eats 2 of the lollipops. With the remaining lollipops, Jean wants to package 2 lollipops in one bag. How many bags can Jean fill?"""  # expecting answering directly
-
+TL_CHAT9 = """What's your name?"""
+TL_CHAT10 = "Hi"
+TL_CHAT11 = "Tell me about your team"
 
 if __name__ == "__main__":
     # NOTE: Add access_token to test github issue fixing