From 78462d5199925fc0080e175f5787e6cf8b96dc93 Mon Sep 17 00:00:00 2001
From: garylin2099 <yzlin@fuzhi.ai>
Date: Mon, 29 Jul 2024 15:07:27 +0800
Subject: [PATCH] hard rule to ask human

---
 metagpt/prompts/di/role_zero.py | 12 ++++++++++++
 metagpt/roles/di/role_zero.py   |  6 ++++++
 2 files changed, 18 insertions(+)

diff --git a/metagpt/prompts/di/role_zero.py b/metagpt/prompts/di/role_zero.py
index 478f5c0d4..de1673e62 100644
--- a/metagpt/prompts/di/role_zero.py
+++ b/metagpt/prompts/di/role_zero.py
@@ -80,6 +80,18 @@ Review and reflect on the history carefully, provide a different response.
 Describe if you should terminate using **end** command, or use **RoleZero.ask_human** to ask human for help, or try a different approach and output different commands. You are NOT allowed to provide the same commands again.
 Your reflection, then the commands in a json array:
 """
+ASK_HUMAN_COMMAND = """
+```json
+[
+    {
+        "command_name": "RoleZero.ask_human",
+        "args": {
+            "question": "I'm a little uncertain about the next step, could you provide me with some guidance?"
+        }
+    }
+]
+```
+"""
 JSON_REPAIR_PROMPT = """
 ## json data
 {json_data}
diff --git a/metagpt/roles/di/role_zero.py b/metagpt/roles/di/role_zero.py
index 6d6eb26ce..ad5c2d26a 100644
--- a/metagpt/roles/di/role_zero.py
+++ b/metagpt/roles/di/role_zero.py
@@ -15,6 +15,7 @@ from metagpt.exp_pool.context_builders import RoleZeroContextBuilder
 from metagpt.exp_pool.serializers import RoleZeroSerializer
 from metagpt.logs import logger
 from metagpt.prompts.di.role_zero import (
+    ASK_HUMAN_COMMAND,
     CMD_PROMPT,
     JSON_REPAIR_PROMPT,
     QUICK_THINK_PROMPT,
@@ -272,6 +273,11 @@ class RoleZero(Role):
             # 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
+
+            #  Hard rule to ask human for help
+            if past_rsp.count(command_rsp) >= 3:
+                return ASK_HUMAN_COMMAND
+            # Try correction by self
             logger.warning(f"Duplicate response detected: {command_rsp}")
             regenerate_req = req + [UserMessage(content=REGENERATE_PROMPT)]
             regenerate_req = self.llm.format_msg(regenerate_req)