diff --git a/src/chat/focus_chat/planners/planner.py b/src/chat/focus_chat/planners/planner.py index 911c98b5..a914e20b 100644 --- a/src/chat/focus_chat/planners/planner.py +++ b/src/chat/focus_chat/planners/planner.py @@ -24,41 +24,39 @@ install(extra_lines=3) def init_prompt(): Prompt( """ -Your self-awareness is: +你的自我认知是： {self_info_block} {extra_info_block} - -You need to decide how to participate in the conversation based on the following information -These information may conflict, please integrate these information, and choose the most suitable action: - +你需要基于以下信息决定如何参与对话 +这些信息可能会有冲突，请你整合这些信息，并选择一个最合适的action： {chat_content_block} {mind_info_block} {cycle_info_block} -IMPORTANT: The following tool call information has the highest priority and should be heavily weighted in your decision-making: -{structured_info_block} - -Please analyze the conversation content and new messages you see, refer to the conversation plan, and choose the appropriate action: +请综合分析聊天内容和你看到的新消息，参考聊天规划，选择合适的action: +注意，除了下面动作选项之外，你在群聊里不能做其他任何事情，这是你能力的边界，现在请你选择合适的action: {action_options_text} -You must choose one from the available actions above and explain why. -Your decision must be output in strict JSON format, and only contain JSON content, no other text or explanation. +你必须从上面列出的可用action中选择一个，并说明原因。 +你的决策必须以严格的 JSON 格式输出，且仅包含 JSON 内容，不要有任何其他文字或解释。 -Please output your decision JSON in the following format: +{moderation_prompt} + +请你以下面格式输出你选择的action： {{ "action": "action_name", - "reasoning": "Your decision reason", - "parameter1": "parameter1 value", - "parameter2": "parameter2 value", - "parameter3": "parameter3 value", + "reasoning": "你的决策理由", + "参数1": "参数1的值", + "参数2": "参数2的值", + "参数3": "参数3的值", ... }} -Please output your decision JSON: """, +请输出你的决策 JSON：""", "planner_prompt", ) @@ -81,7 +79,7 @@ class ActionPlanner: self.planner_llm = LLMRequest( model=global_config.model.focus_planner, max_tokens=1000, - request_type="action_planning", # 用于动作规划 + request_type="focus_planner", # 用于动作规划 ) self.action_manager = action_manager @@ -98,20 +96,10 @@ class ActionPlanner: action = "no_reply" # 默认动作 reasoning = "规划器初始化默认" action_data = {} - - # 初始化所有可能用到的变量，避免UnboundLocalError - observed_messages = [] - observed_messages_str = "" - current_mind = "" - cycle_info = "" - self_info = "" - is_group_chat = True # 默认为群聊 - prompt = "未生成prompt" # 初始化prompt变量 try: # 获取观察信息 extra_info: list[str] = [] - _structured_info_list = [] # 首先处理动作变更 for info in all_plan_info: @@ -119,6 +107,7 @@ class ActionPlanner: add_actions = info.get_add_actions() remove_actions = info.get_remove_actions() reason = info.get_reason() + print(f"{self.log_prefix} 动作变更: {add_actions} {remove_actions} {reason}") # 处理动作的增加 for action_name in add_actions: @@ -137,6 +126,10 @@ class ActionPlanner: reasoning = f"之前选择的动作{action}已被移除，原因: {reason}" # 继续处理其他信息 + self_info = "" + current_mind = "" + cycle_info = "" + structured_info = "" for info in all_plan_info: if isinstance(info, ObsInfo): observed_messages = info.get_talking_message() @@ -150,24 +143,25 @@ class ActionPlanner: elif isinstance(info, SelfInfo): self_info = info.get_processed_info() elif isinstance(info, StructuredInfo): - _structured_info = info.get_data() - # 收集工具调用的结构化信息 - if _structured_info and isinstance(_structured_info, dict): - # StructuredInfo 的数据结构是 {tool_type: tool_content} - for tool_type, tool_content in _structured_info.items(): - if tool_content: # 确保内容不为空 - _structured_info_list.append(f"{tool_type}: {str(tool_content)}") + structured_info = info.get_processed_info() + # print(f"structured_info: {structured_info}") elif not isinstance(info, ActionInfo): # 跳过已处理的ActionInfo extra_info.append(info.get_processed_info()) # 获取当前可用的动作 current_available_actions = self.action_manager.get_using_actions() - # 如果没有可用动作，直接返回no_reply - if not current_available_actions: - logger.warning(f"{self.log_prefix}没有可用的动作，将使用no_reply") + # 如果没有可用动作或只有no_reply动作，直接返回no_reply + if not current_available_actions or ( + len(current_available_actions) == 1 and "no_reply" in current_available_actions + ): action = "no_reply" - reasoning = "没有可用的动作" + reasoning = "没有可用的动作" if not current_available_actions else "只有no_reply动作可用，跳过规划" + logger.info(f"{self.log_prefix}{reasoning}") + self.action_manager.restore_actions() + logger.debug( + f"{self.log_prefix}沉默后恢复到默认动作集, 当前可用: {list(self.action_manager.get_using_actions().keys())}" + ) return { "action_result": {"action_type": action, "action_data": action_data, "reasoning": reasoning}, "current_mind": current_mind, @@ -175,20 +169,13 @@ class ActionPlanner: } # --- 构建提示词 (调用修改后的 PromptBuilder 方法) --- - # 构建工具信息块 - structured_info_block_str = "\n".join(_structured_info_list) - if structured_info_block_str: - structured_info_block_str = f"The following is basic information returned by tool calls. Please use this information as the basis for subsequent decision-making and actions:\n{structured_info_block_str}" - else: - structured_info_block_str = "No tool information available for reference." - prompt = await self.build_planner_prompt( self_info_block=self_info, is_group_chat=is_group_chat, # <-- Pass HFC state chat_target_info=None, observed_messages_str=observed_messages_str, # <-- Pass local variable current_mind=current_mind, # <-- Pass argument - structured_info_block=structured_info_block_str, + structured_info=structured_info, # <-- Pass SubMind info current_available_actions=current_available_actions, # <-- Pass determined actions cycle_info=cycle_info, # <-- Pass cycle info extra_info=extra_info, @@ -197,8 +184,9 @@ class ActionPlanner: # --- 调用 LLM (普通文本生成) --- llm_content = None try: - llm_content, _, _ = await self.planner_llm.generate_response(prompt=prompt) + llm_content, reasoning_content, _ = await self.planner_llm.generate_response(prompt=prompt) logger.debug(f"{self.log_prefix}[Planner] LLM 原始 JSON 响应 (预期): {llm_content}") + logger.debug(f"{self.log_prefix}[Planner] LLM 原始理由 响应 (预期): {reasoning_content}") except Exception as req_e: logger.error(f"{self.log_prefix}[Planner] LLM 请求执行失败: {req_e}") reasoning = f"LLM 请求失败，你的模型出现问题: {req_e}" @@ -254,10 +242,10 @@ class ActionPlanner: f"{self.log_prefix}规划器Prompt:\n{prompt}\n\n决策动作:{action},\n动作信息: '{action_data}'\n理由: {reasoning}" ) - # 恢复原始动作集 + # 恢复到默认动作集 self.action_manager.restore_actions() logger.debug( - f"{self.log_prefix}恢复了原始动作集, 当前可用: {list(self.action_manager.get_using_actions().keys())}" + f"{self.log_prefix}规划后恢复到默认动作集, 当前可用: {list(self.action_manager.get_using_actions().keys())}" ) action_result = {"action_type": action, "action_data": action_data, "reasoning": reasoning} @@ -277,7 +265,7 @@ class ActionPlanner: chat_target_info: Optional[dict], # Now passed as argument observed_messages_str: str, current_mind: Optional[str], - structured_info_block: str, + structured_info: Optional[str], current_available_actions: Dict[str, ActionInfo], cycle_info: Optional[str], extra_info: list[str], @@ -334,11 +322,14 @@ class ActionPlanner: action_options_block += using_action_prompt - extra_info_block = "\n".join(extra_info) # 将局部变量名从 extra_info_block_str 改为 extra_info_block - if extra_info_block: # 使用新的变量名进行检查 - extra_info_block = f"The following is some additional information. Please read the following content to make a decision:\n{extra_info_block}\nEnd of additional information." # 使用新的变量名进行格式化 + extra_info_block = "\n".join(extra_info) + extra_info_block += f"\n{structured_info}" + if extra_info or structured_info: + extra_info_block = f"以下是一些额外的信息，现在请你阅读以下内容，进行决策\n{extra_info_block}\n以上是一些额外的信息，现在请你阅读以下内容，进行决策" else: - extra_info_block = "No additional information available." # 使用新的变量名进行赋值 + extra_info_block = "" + + moderation_prompt_block = "请不要输出违法违规内容，不要输出色情，暴力，政治相关内容，如有敏感内容，请规避。" planner_prompt_template = await global_prompt_manager.get_prompt_async("planner_prompt") prompt = planner_prompt_template.format( @@ -348,10 +339,11 @@ class ActionPlanner: chat_context_description=chat_context_description, chat_content_block=chat_content_block, mind_info_block=mind_info_block, - structured_info_block=structured_info_block, cycle_info_block=cycle_info, action_options_text=action_options_block, - extra_info_block=extra_info_block, # 确保这里传递的是修改后的变量名 + # action_available_block=action_available_block, + extra_info_block=extra_info_block, + moderation_prompt=moderation_prompt_block, ) return prompt