feat: 重构Kokoro Flow Chatter，新增规划器和回复生成器，优化提示词构建逻辑

src/plugins/built_in/kokoro_flow_chatter/__init__.py

@@ -19,7 +19,8 @@ from .models import (
 )
 from .session import KokoroSession, SessionManager, get_session_manager
 from .chatter import KokoroFlowChatter
-from .replyer import generate_response
+from .planner import generate_plan
+from .replyer import generate_reply_text
 from .proactive_thinker import (
     ProactiveThinker,
     get_proactive_thinker,
@@ -60,7 +61,8 @@ __all__ = [
     "get_session_manager",
     # Core Components
     "KokoroFlowChatter",
-    "generate_response",
+    "generate_plan",
+    "generate_reply_text",
     # Proactive Thinker
     "ProactiveThinker",
     "get_proactive_thinker",

src/plugins/built_in/kokoro_flow_chatter/chatter.py

@@ -19,7 +19,8 @@ from src.plugin_system.base.base_chatter import BaseChatter
 from src.plugin_system.base.component_types import ChatType
 
 from .models import SessionStatus
-from .replyer import generate_response
+from .planner import generate_plan
+from .replyer import generate_reply_text
 from .session import get_session_manager
 
 if TYPE_CHECKING:
@@ -143,8 +144,8 @@ class KokoroFlowChatter(BaseChatter):
                 # 8. 获取聊天流
                 chat_stream = await self._get_chat_stream()
                 
-                # 9. 调用 Replyer 生成响应
+                # 9. 调用 Planner 生成行动计划
-                response = await generate_response(
+                plan_response = await generate_plan(
                     session=session,
                     user_name=user_name,
                     situation_type=situation_type,
@@ -152,15 +153,35 @@ class KokoroFlowChatter(BaseChatter):
                     available_actions=available_actions,
                 )
                 
-                # 10. 执行动作作
+                # 10. 对于需要回复的动作，调用 Replyer 生成实际文本
+                processed_actions = []
+                for action in plan_response.actions:
+                    if action.type == "kfc_reply":
+                        # 调用 replyer 生成回复文本
+                        success, reply_text = await generate_reply_text(
+                            session=session,
+                            user_name=user_name,
+                            thought=plan_response.thought,
+                            situation_type=situation_type,
+                            chat_stream=chat_stream,
+                        )
+                        if success and reply_text:
+                            # 更新 action 的 content
+                            action.params["content"] = reply_text
+                        else:
+                            logger.warning("[KFC] 回复生成失败，跳过该动作")
+                            continue
+                    processed_actions.append(action)
+                
+                # 11. 执行动作
                 exec_results = []
                 has_reply = False
-                for action in response.actions:
+                for action in processed_actions:
                     result = await self.action_manager.execute_action(
                         action_name=action.type,
                         chat_id=self.stream_id,
                         target_message=target_message,
-                        reasoning=response.thought,
+                        reasoning=plan_response.thought,
                         action_data=action.params,
                         thinking_id=None,
                         log_prefix="[KFC]",
@@ -169,31 +190,31 @@ class KokoroFlowChatter(BaseChatter):
                     if result.get("success") and action.type in ("kfc_reply", "respond"):
                         has_reply = True
                 
-                # 11. 记录 Bot 规划到 mental_log
+                # 12. 记录 Bot 规划到 mental_log
                 session.add_bot_planning(
-                    thought=response.thought,
+                    thought=plan_response.thought,
-                    actions=[a.to_dict() for a in response.actions],
+                    actions=[a.to_dict() for a in processed_actions],
-                    expected_reaction=response.expected_reaction,
+                    expected_reaction=plan_response.expected_reaction,
-                    max_wait_seconds=response.max_wait_seconds,
+                    max_wait_seconds=plan_response.max_wait_seconds,
                 )
                 
-                # 12. 更新 Session 状态
+                # 13. 更新 Session 状态
-                if response.max_wait_seconds > 0:
+                if plan_response.max_wait_seconds > 0:
                     session.start_waiting(
-                        expected_reaction=response.expected_reaction,
+                        expected_reaction=plan_response.expected_reaction,
-                        max_wait_seconds=response.max_wait_seconds,
+                        max_wait_seconds=plan_response.max_wait_seconds,
                     )
                 else:
                     session.end_waiting()
                 
-                # 13. 标记消息为已读
+                # 14. 标记消息为已读
                 for msg in unread_messages:
                     context.mark_message_as_read(str(msg.message_id))
                 
-                # 14. 保存 Session
+                # 15. 保存 Session
                 await self.session_manager.save_session(user_id)
                 
-                # 15. 更新统计
+                # 16. 更新统计
                 self._stats["messages_processed"] += len(unread_messages)
                 if has_reply:
                     self._stats["successful_responses"] += 1
@@ -201,15 +222,15 @@ class KokoroFlowChatter(BaseChatter):
                 logger.info(
                     f"{SOFT_PURPLE}[KFC]{RESET} 处理完成: "
                     f"user={user_name}, situation={situation_type}, "
-                    f"actions={[a.type for a in response.actions]}, "
+                    f"actions={[a.type for a in processed_actions]}, "
-                    f"wait={response.max_wait_seconds}s"
+                    f"wait={plan_response.max_wait_seconds}s"
                 )
                 
                 return self._build_result(
                     success=True,
                     message="processed",
                     has_reply=has_reply,
-                    thought=response.thought,
+                    thought=plan_response.thought,
                     situation_type=situation_type,
                 )
                 

src/plugins/built_in/kokoro_flow_chatter/planner.py

@@ -0,0 +1,112 @@
+"""
+Kokoro Flow Chatter - Planner
+
+规划器：负责分析情境并生成行动计划
+- 输入：会话状态、用户消息、情境类型
+- 输出：LLMResponse（包含 thought、actions、expected_reaction、max_wait_seconds）
+- 不负责生成具体回复文本，只决定"要做什么"
+"""
+
+from typing import TYPE_CHECKING, Optional
+
+from src.common.logger import get_logger
+from src.plugin_system.apis import llm_api
+from src.utils.json_parser import extract_and_parse_json
+
+from .models import LLMResponse
+from .prompt.builder import get_prompt_builder
+from .session import KokoroSession
+
+if TYPE_CHECKING:
+    from src.chat.message_receive.chat_stream import ChatStream
+
+logger = get_logger("kfc_planner")
+
+
+async def generate_plan(
+    session: KokoroSession,
+    user_name: str,
+    situation_type: str = "new_message",
+    chat_stream: Optional["ChatStream"] = None,
+    available_actions: Optional[dict] = None,
+    extra_context: Optional[dict] = None,
+) -> LLMResponse:
+    """
+    生成行动计划
+    
+    Args:
+        session: 会话对象
+        user_name: 用户名称
+        situation_type: 情况类型
+        chat_stream: 聊天流对象
+        available_actions: 可用动作字典
+        extra_context: 额外上下文
+        
+    Returns:
+        LLMResponse 对象，包含计划信息
+    """
+    try:
+        # 1. 构建规划器提示词
+        prompt_builder = get_prompt_builder()
+        prompt = await prompt_builder.build_planner_prompt(
+            session=session,
+            user_name=user_name,
+            situation_type=situation_type,
+            chat_stream=chat_stream,
+            available_actions=available_actions,
+            extra_context=extra_context,
+        )
+        
+        from src.config.config import global_config
+        if global_config and global_config.debug.show_prompt:
+            logger.info(f"[KFC Planner] 生成的规划提示词:\n{prompt}")
+        
+        # 2. 获取 planner 模型配置并调用 LLM
+        models = llm_api.get_available_models()
+        planner_config = models.get("planner")
+        
+        if not planner_config:
+            logger.error("[KFC Planner] 未找到 planner 模型配置")
+            return LLMResponse.create_error_response("未找到 planner 模型配置")
+        
+        success, raw_response, reasoning, model_name = await llm_api.generate_with_model(
+            prompt=prompt,
+            model_config=planner_config,
+            request_type="kokoro_flow_chatter.plan",
+        )
+        
+        if not success:
+            logger.error(f"[KFC Planner] LLM 调用失败: {raw_response}")
+            return LLMResponse.create_error_response(raw_response)
+        
+        logger.debug(f"[KFC Planner] LLM 响应 (model={model_name}):\n{raw_response}")
+        
+        # 3. 解析响应
+        return _parse_response(raw_response)
+        
+    except Exception as e:
+        logger.error(f"[KFC Planner] 生成失败: {e}")
+        import traceback
+        traceback.print_exc()
+        return LLMResponse.create_error_response(str(e))
+
+
+def _parse_response(raw_response: str) -> LLMResponse:
+    """解析 LLM 响应"""
+    data = extract_and_parse_json(raw_response, strict=False)
+    
+    if not data or not isinstance(data, dict):
+        logger.warning(f"[KFC Planner] 无法解析 JSON: {raw_response[:200]}...")
+        return LLMResponse.create_error_response("无法解析响应格式")
+    
+    response = LLMResponse.from_dict(data)
+    
+    if response.thought:
+        logger.info(
+            f"[KFC Planner] 解析成功: thought={response.thought[:50]}..., "
+            f"actions={[a.type for a in response.actions]}"
+        )
+    else:
+        logger.warning("[KFC Planner] 响应缺少 thought")
+    
+    return response

src/plugins/built_in/kokoro_flow_chatter/proactive_thinker.py

@@ -21,7 +21,8 @@ from src.config.config import global_config
 from src.plugin_system.apis.unified_scheduler import TriggerType, unified_scheduler
 
 from .models import EventType, SessionStatus
-from .replyer import generate_response
+from .planner import generate_plan
+from .replyer import generate_reply_text
 from .session import KokoroSession, get_session_manager
 
 if TYPE_CHECKING:
@@ -288,8 +289,8 @@ class ProactiveThinker:
             action_modifier = ActionModifier(action_manager, session.stream_id)
             await action_modifier.modify_actions(chatter_name="KokoroFlowChatter")
             
-            # 调用 Replyer 生成超时决策
+            # 调用 Planner 生成超时决策
-            response = await generate_response(
+            plan_response = await generate_plan(
                 session=session,
                 user_name=session.user_id,  # 这里可以改进，获取真实用户名
                 situation_type="timeout",
@@ -297,32 +298,50 @@ class ProactiveThinker:
                 available_actions=action_manager.get_using_actions(),
             )
             
+            # 对于需要回复的动作，调用 Replyer 生成实际文本
+            processed_actions = []
+            for action in plan_response.actions:
+                if action.type == "kfc_reply":
+                    success, reply_text = await generate_reply_text(
+                        session=session,
+                        user_name=session.user_id,
+                        thought=plan_response.thought,
+                        situation_type="timeout",
+                        chat_stream=chat_stream,
+                    )
+                    if success and reply_text:
+                        action.params["content"] = reply_text
+                    else:
+                        logger.warning("[ProactiveThinker] 回复生成失败，跳过该动作")
+                        continue
+                processed_actions.append(action)
+            
             # 执行动作
-            for action in response.actions:
+            for action in processed_actions:
                 await action_manager.execute_action(
                     action_name=action.type,
                     chat_id=session.stream_id,
                     target_message=None,
-                    reasoning=response.thought,
+                    reasoning=plan_response.thought,
                     action_data=action.params,
                     thinking_id=None,
-                    log_prefix="[KFC V2 ProactiveThinker]",
+                    log_prefix="[KFC ProactiveThinker]",
                 )
             
             # 记录到 mental_log
             session.add_bot_planning(
-                thought=response.thought,
+                thought=plan_response.thought,
-                actions=[a.to_dict() for a in response.actions],
+                actions=[a.to_dict() for a in processed_actions],
-                expected_reaction=response.expected_reaction,
+                expected_reaction=plan_response.expected_reaction,
-                max_wait_seconds=response.max_wait_seconds,
+                max_wait_seconds=plan_response.max_wait_seconds,
             )
             
             # 更新状态
-            if response.max_wait_seconds > 0:
+            if plan_response.max_wait_seconds > 0:
                 # 继续等待
                 session.start_waiting(
-                    expected_reaction=response.expected_reaction,
+                    expected_reaction=plan_response.expected_reaction,
-                    max_wait_seconds=response.max_wait_seconds,
+                    max_wait_seconds=plan_response.max_wait_seconds,
                 )
             else:
                 # 不再等待
@@ -333,8 +352,8 @@ class ProactiveThinker:
             
             logger.info(
                 f"[ProactiveThinker] 超时决策完成: user={session.user_id}, "
-                f"actions={[a.type for a in response.actions]}, "
+                f"actions={[a.type for a in processed_actions]}, "
-                f"continue_wait={response.max_wait_seconds > 0}"
+                f"continue_wait={plan_response.max_wait_seconds > 0}"
             )
             
         except Exception as e:
@@ -449,23 +468,25 @@ class ProactiveThinker:
             else:
                 silence_duration = f"{silence_seconds / 3600:.1f} 小时"
             
-            # 调用 Replyer
+            extra_context = {
-            response = await generate_response(
+                "trigger_reason": trigger_reason,
+                "silence_duration": silence_duration,
+            }
+            
+            # 调用 Planner
+            plan_response = await generate_plan(
                 session=session,
                 user_name=session.user_id,
                 situation_type="proactive",
                 chat_stream=chat_stream,
                 available_actions=action_manager.get_using_actions(),
-                extra_context={
+                extra_context=extra_context,
-                    "trigger_reason": trigger_reason,
-                    "silence_duration": silence_duration,
-                },
             )
             
             # 检查是否决定不打扰
             is_do_nothing = (
-                len(response.actions) == 0 or
+                len(plan_response.actions) == 0 or
-                (len(response.actions) == 1 and response.actions[0].type == "do_nothing")
+                (len(plan_response.actions) == 1 and plan_response.actions[0].type == "do_nothing")
             )
             
             if is_do_nothing:
@@ -474,32 +495,51 @@ class ProactiveThinker:
                 await self.session_manager.save_session(session.user_id)
                 return
             
+            # 对于需要回复的动作，调用 Replyer 生成实际文本
+            processed_actions = []
+            for action in plan_response.actions:
+                if action.type == "kfc_reply":
+                    success, reply_text = await generate_reply_text(
+                        session=session,
+                        user_name=session.user_id,
+                        thought=plan_response.thought,
+                        situation_type="proactive",
+                        chat_stream=chat_stream,
+                        extra_context=extra_context,
+                    )
+                    if success and reply_text:
+                        action.params["content"] = reply_text
+                    else:
+                        logger.warning("[ProactiveThinker] 回复生成失败，跳过该动作")
+                        continue
+                processed_actions.append(action)
+            
             # 执行动作
-            for action in response.actions:
+            for action in processed_actions:
                 await action_manager.execute_action(
                     action_name=action.type,
                     chat_id=session.stream_id,
                     target_message=None,
-                    reasoning=response.thought,
+                    reasoning=plan_response.thought,
                     action_data=action.params,
                     thinking_id=None,
-                    log_prefix="[KFC V2 ProactiveThinker]",
+                    log_prefix="[KFC ProactiveThinker]",
                 )
             
             # 记录到 mental_log
             session.add_bot_planning(
-                thought=response.thought,
+                thought=plan_response.thought,
-                actions=[a.to_dict() for a in response.actions],
+                actions=[a.to_dict() for a in processed_actions],
-                expected_reaction=response.expected_reaction,
+                expected_reaction=plan_response.expected_reaction,
-                max_wait_seconds=response.max_wait_seconds,
+                max_wait_seconds=plan_response.max_wait_seconds,
             )
             
             # 更新状态
             session.last_proactive_at = time.time()
-            if response.max_wait_seconds > 0:
+            if plan_response.max_wait_seconds > 0:
                 session.start_waiting(
-                    expected_reaction=response.expected_reaction,
+                    expected_reaction=plan_response.expected_reaction,
-                    max_wait_seconds=response.max_wait_seconds,
+                    max_wait_seconds=plan_response.max_wait_seconds,
                 )
             
             # 保存
@@ -507,7 +547,7 @@ class ProactiveThinker:
             
             logger.info(
                 f"[ProactiveThinker] 主动发起完成: user={session.user_id}, "
-                f"actions={[a.type for a in response.actions]}"
+                f"actions={[a.type for a in processed_actions]}"
             )
             
         except Exception as e:

src/plugins/built_in/kokoro_flow_chatter/prompt/builder.py

@@ -38,7 +38,7 @@ class PromptBuilder:
     def __init__(self):
         self._context_builder = None
     
-    async def build_prompt(
+    async def build_planner_prompt(
         self,
         session: KokoroSession,
         user_name: str,
@@ -48,7 +48,7 @@ class PromptBuilder:
         extra_context: Optional[dict] = None,
     ) -> str:
         """
-        构建完整的提示词
+        构建规划器提示词（用于生成行动计划）
         
         Args:
             session: 会话对象
@@ -59,7 +59,7 @@ class PromptBuilder:
             extra_context: 额外上下文（如 trigger_reason）
             
         Returns:
-            完整的提示词
+            完整的规划器提示词
         """
         extra_context = extra_context or {}
         
@@ -89,8 +89,8 @@ class PromptBuilder:
         # 6. 构建可用动作
         actions_block = self._build_actions_block(available_actions)
         
-        # 7. 获取输出格式
+        # 7. 获取规划器输出格式
-        output_format = await self._get_output_format()
+        output_format = await self._get_planner_output_format()
         
         # 8. 使用统一的 prompt 管理系统格式化
         prompt = await global_prompt_manager.format_prompt(
@@ -109,6 +109,76 @@ class PromptBuilder:
         
         return prompt
     
+    async def build_replyer_prompt(
+        self,
+        session: KokoroSession,
+        user_name: str,
+        thought: str,
+        situation_type: str = "new_message",
+        chat_stream: Optional["ChatStream"] = None,
+        extra_context: Optional[dict] = None,
+    ) -> str:
+        """
+        构建回复器提示词（用于生成自然的回复文本）
+        
+        Args:
+            session: 会话对象
+            user_name: 用户名称
+            thought: 规划器生成的想法
+            situation_type: 情况类型
+            chat_stream: 聊天流对象
+            extra_context: 额外上下文
+            
+        Returns:
+            完整的回复器提示词
+        """
+        extra_context = extra_context or {}
+        
+        # 获取 user_id
+        user_id = session.user_id if session else None
+        
+        # 1. 构建人设块
+        persona_block = self._build_persona_block()
+        
+        # 2. 使用 context_builder 获取关系、记忆、表达习惯等
+        context_data = await self._build_context_data(user_name, chat_stream, user_id)
+        relation_block = context_data.get("relation_info", f"你与 {user_name} 还不太熟悉，这是早期的交流阶段。")
+        memory_block = context_data.get("memory_block", "")
+        expression_habits = self._build_combined_expression_block(context_data.get("expression_habits", ""))
+        
+        # 3. 构建活动流
+        activity_stream = await self._build_activity_stream(session, user_name)
+        
+        # 4. 构建当前情况（简化版，不需要那么详细）
+        current_situation = await self._build_current_situation(
+            session, user_name, situation_type, extra_context
+        )
+        
+        # 5. 构建聊天历史总览
+        chat_history_block = await self._build_chat_history_block(chat_stream)
+        
+        # 6. 构建回复情景上下文
+        reply_context = await self._build_reply_context(
+            session, user_name, situation_type, extra_context
+        )
+        
+        # 7. 使用回复器专用模板
+        prompt = await global_prompt_manager.format_prompt(
+            PROMPT_NAMES["replyer"],
+            user_name=user_name,
+            persona_block=persona_block,
+            relation_block=relation_block,
+            memory_block=memory_block or "（暂无相关记忆）",
+            activity_stream=activity_stream or "（这是你们第一次聊天）",
+            current_situation=current_situation,
+            chat_history_block=chat_history_block,
+            expression_habits=expression_habits or "（根据自然对话风格回复即可）",
+            thought=thought,
+            reply_context=reply_context,
+        )
+        
+        return prompt
+    
     def _build_persona_block(self) -> str:
         """构建人设块"""
         if global_config is None:
@@ -578,6 +648,91 @@ class PromptBuilder:
     "expected_reaction": "期待的反应",
     "max_wait_seconds": 300
 }"""
+    
+    async def _get_planner_output_format(self) -> str:
+        """获取规划器输出格式模板"""
+        try:
+            prompt = await global_prompt_manager.get_prompt_async(
+                PROMPT_NAMES["planner_output_format"]
+            )
+            return prompt.template
+        except KeyError:
+            # 如果模板未注册，返回默认格式
+            return """请用 JSON 格式回复：
+{
+    "thought": "你的想法",
+    "actions": [{"type": "kfc_reply"}],
+    "expected_reaction": "期待的反应",
+    "max_wait_seconds": 300
+}
+
+注意：kfc_reply 动作不需要填写 content 字段，回复内容会单独生成。"""
+    
+    async def _build_reply_context(
+        self,
+        session: KokoroSession,
+        user_name: str,
+        situation_type: str,
+        extra_context: dict,
+    ) -> str:
+        """
+        构建回复情景上下文
+        
+        根据 situation_type 构建不同的情景描述，帮助回复器理解当前要回复的情境
+        """
+        # 获取最后一条用户消息
+        target_message = ""
+        entries = session.get_recent_entries(limit=10)
+        for entry in reversed(entries):
+            if entry.event_type == EventType.USER_MESSAGE:
+                target_message = entry.content or ""
+                break
+        
+        if situation_type == "new_message":
+            return await global_prompt_manager.format_prompt(
+                PROMPT_NAMES["replyer_context_normal"],
+                user_name=user_name,
+                target_message=target_message or "（无消息内容）",
+            )
+        
+        elif situation_type == "reply_in_time":
+            elapsed = session.waiting_config.get_elapsed_seconds()
+            max_wait = session.waiting_config.max_wait_seconds
+            return await global_prompt_manager.format_prompt(
+                PROMPT_NAMES["replyer_context_in_time"],
+                user_name=user_name,
+                target_message=target_message or "（无消息内容）",
+                elapsed_minutes=elapsed / 60,
+                max_wait_minutes=max_wait / 60,
+            )
+        
+        elif situation_type == "reply_late":
+            elapsed = session.waiting_config.get_elapsed_seconds()
+            max_wait = session.waiting_config.max_wait_seconds
+            return await global_prompt_manager.format_prompt(
+                PROMPT_NAMES["replyer_context_late"],
+                user_name=user_name,
+                target_message=target_message or "（无消息内容）",
+                elapsed_minutes=elapsed / 60,
+                max_wait_minutes=max_wait / 60,
+            )
+        
+        elif situation_type == "proactive":
+            silence = extra_context.get("silence_duration", "一段时间")
+            trigger_reason = extra_context.get("trigger_reason", "")
+            return await global_prompt_manager.format_prompt(
+                PROMPT_NAMES["replyer_context_proactive"],
+                user_name=user_name,
+                silence_duration=silence,
+                trigger_reason=trigger_reason,
+            )
+        
+        # 默认使用普通情景
+        return await global_prompt_manager.format_prompt(
+            PROMPT_NAMES["replyer_context_normal"],
+            user_name=user_name,
+            target_message=target_message or "（无消息内容）",
+        )
 
 
 # 全局单例

src/plugins/built_in/kokoro_flow_chatter/prompt/prompts.py

@@ -212,10 +212,141 @@ kfc_ENTRY_PROACTIVE_TRIGGER = Prompt(
 """,
 )
 
+# =================================================================================================
+# Planner 专用输出格式
+# =================================================================================================
+
+kfc_PLANNER_OUTPUT_FORMAT = Prompt(
+    name="kfc_planner_output_format",
+    template="""请用以下 JSON 格式回复：
+```json
+{{
+    "thought": "你脑子里在想什么，越自然越好",
+    "actions": [
+        {{"type": "动作名称", ...动作参数}}
+    ],
+    "expected_reaction": "你期待对方的反应是什么",
+    "max_wait_seconds": 300
+}}
+```
+
+### 字段说明
+- `thought`：你的内心独白，记录你此刻的想法和感受。要自然，不要技术性语言。
+- `actions`：你要执行的动作列表。每个动作是一个对象，必须包含 `type` 字段指定动作类型，其他字段根据动作类型不同而不同（参考上面每个动作的示例）。
+  - 对于 `kfc_reply` 动作，只需要指定 `{{"type": "kfc_reply"}}`，不需要填写 `content` 字段（回复内容会单独生成）
+- `expected_reaction`：你期待对方如何回应（用于判断是否需要等待）
+- `max_wait_seconds`：设定等待时间（秒），0 表示不等待，超时后你会考虑是否要主动说点什么
+
+### 注意事项
+- 动作参数直接写在动作对象里，不需要 `action_data` 包装
+- 即使什么都不想做，也放一个 `{{"type": "do_nothing"}}`
+- 可以组合多个动作，比如先发消息再发表情""",
+)
+
+# =================================================================================================
+# Replyer 专用提示词模板
+# =================================================================================================
+
+kfc_REPLYER_PROMPT = Prompt(
+    name="kfc_replyer",
+    template="""# 你与 {user_name} 的私聊
+
+## 1. 你是谁
+{persona_block}
+
+## 2. 你与 {user_name} 的关系
+{relation_block}
+
+## 3. 相关记忆
+{memory_block}
+
+## 4. 你们之间发生的事（活动流）
+以下是你和 {user_name} 最近的互动历史，按时间顺序记录了你们的对话和你的心理活动：
+
+{activity_stream}
+
+## 5. 当前情况
+{current_situation}
+
+## 6. 聊天历史总览
+以下是你和 {user_name} 的聊天记录，帮助你更好地理解对话上下文：
+
+{chat_history_block}
+
+## 7. 你的表达习惯
+{expression_habits}
+
+## 8. 你的决策
+你已经决定要回复 {user_name}。
+
+**你的想法**：{thought}
+
+{reply_context}
+
+## 9. 回复要求
+
+**情景化表达**：
+- 根据你们的关系亲密度选择合适的语气和称呼
+- 参考活动流中的互动历史，保持对话的连贯性
+- 如果对方回复得晚/早，可以自然地表现出你的感受
+
+**自然对话**：
+- 像真实的朋友聊天一样，不要生硬或公式化
+- 可以用口语、网络用语、语气词，让回复更生动
+- 长度适中，不要太长也不要太短
+
+**表达习惯**：
+- 参考上面的"表达习惯"部分，使用你习惯的语言风格
+- 保持人设的一致性
+
+**禁忌**：
+- 不要重复你之前说过的话
+- 不要输出 JSON 格式或技术性语言
+- 不要加引号、括号等多余符号
+- 不要用"我决定..."、"因此..."这种总结性语言
+
+现在，请直接输出你要说的话：""",
+)
+
+kfc_REPLYER_CONTEXT_NORMAL = Prompt(
+    name="kfc_replyer_context_normal",
+    template="""你要回复的是 {user_name} 刚发来的消息：
+「{target_message}」""",
+)
+
+kfc_REPLYER_CONTEXT_IN_TIME = Prompt(
+    name="kfc_replyer_context_in_time",
+    template="""你等了 {elapsed_minutes:.1f} 分钟（原本打算最多等 {max_wait_minutes:.1f} 分钟），{user_name} 终于回复了：
+「{target_message}」
+
+你可以表现出一点"等到了回复"的欣喜或轻松。""",
+)
+
+kfc_REPLYER_CONTEXT_LATE = Prompt(
+    name="kfc_replyer_context_late",
+    template="""你等了 {elapsed_minutes:.1f} 分钟（原本只打算等 {max_wait_minutes:.1f} 分钟），{user_name} 才回复：
+「{target_message}」
+
+虽然有点晚，但对方终于回复了。你可以选择轻轻抱怨一下，也可以装作没在意。""",
+)
+
+kfc_REPLYER_CONTEXT_PROACTIVE = Prompt(
+    name="kfc_replyer_context_proactive",
+    template="""你们已经有一段时间（{silence_duration}）没聊天了。{trigger_reason}
+
+你决定主动打破沉默，找 {user_name} 聊点什么。想一个自然的开场白，不要太突兀。""",
+)
+
 # 导出所有模板名称，方便外部引用
 PROMPT_NAMES = {
     "main": "kfc_main",
     "output_format": "kfc_output_format",
+    "planner_output_format": "kfc_planner_output_format",
+    "replyer": "kfc_replyer",
+    "replyer_context_normal": "kfc_replyer_context_normal",
+    "replyer_context_in_time": "kfc_replyer_context_in_time",
+    "replyer_context_late": "kfc_replyer_context_late",
+    "replyer_context_proactive": "kfc_replyer_context_proactive",
     "situation_new_message": "kfc_situation_new_message",
     "situation_reply_in_time": "kfc_situation_reply_in_time",
     "situation_reply_late": "kfc_situation_reply_late",

src/plugins/built_in/kokoro_flow_chatter/replyer.py

@@ -1,16 +1,17 @@
 """
 Kokoro Flow Chatter - Replyer
 
-简化的回复生成模块，使用插件系统的 llm_api
+纯粹的回复生成器：
+- 接收 planner 的决策（thought 等）
+- 专门负责将回复意图转化为自然的对话文本
+- 不输出 JSON，直接生成可发送的消息文本
 """
 
 from typing import TYPE_CHECKING, Optional
 
 from src.common.logger import get_logger
 from src.plugin_system.apis import llm_api
-from src.utils.json_parser import extract_and_parse_json
 
-from .models import LLMResponse
 from .prompt.builder import get_prompt_builder
 from .session import KokoroSession
 
@@ -20,90 +21,103 @@ if TYPE_CHECKING:
 logger = get_logger("kfc_replyer")
 
 
-async def generate_response(
+async def generate_reply_text(
     session: KokoroSession,
     user_name: str,
+    thought: str,
     situation_type: str = "new_message",
     chat_stream: Optional["ChatStream"] = None,
-    available_actions: Optional[dict] = None,
     extra_context: Optional[dict] = None,
-) -> LLMResponse:
+) -> tuple[bool, str]:
     """
-    生成回复
+    生成回复文本
     
     Args:
         session: 会话对象
         user_name: 用户名称
+        thought: 规划器生成的想法（内心独白）
         situation_type: 情况类型
         chat_stream: 聊天流对象
-        available_actions: 可用动作字典
         extra_context: 额外上下文
         
     Returns:
-        LLMResponse 对象
+        (success, reply_text) 元组
+        - success: 是否成功生成
+        - reply_text: 生成的回复文本
     """
     try:
-        # 1. 构建提示词
+        # 1. 构建回复器提示词
         prompt_builder = get_prompt_builder()
-        prompt = await prompt_builder.build_prompt(
+        prompt = await prompt_builder.build_replyer_prompt(
             session=session,
             user_name=user_name,
+            thought=thought,
             situation_type=situation_type,
             chat_stream=chat_stream,
-            available_actions=available_actions,
             extra_context=extra_context,
         )
         
         from src.config.config import global_config
         if global_config and global_config.debug.show_prompt:
-            logger.info(f"[KFC Replyer] 生成的提示词:\n{prompt}")
+            logger.info(f"[KFC Replyer] 生成的回复提示词:\n{prompt}")
         
-        # 2. 获取模型配置并调用 LLM
+        # 2. 获取 replyer 模型配置并调用 LLM
         models = llm_api.get_available_models()
         replyer_config = models.get("replyer")
         
         if not replyer_config:
             logger.error("[KFC Replyer] 未找到 replyer 模型配置")
-            return LLMResponse.create_error_response("未找到 replyer 模型配置")
+            return False, "（回复生成失败：未找到模型配置）"
         
         success, raw_response, reasoning, model_name = await llm_api.generate_with_model(
             prompt=prompt,
             model_config=replyer_config,
-            request_type="kokoro_flow_chatter",
+            request_type="kokoro_flow_chatter.reply",
         )
         
         if not success:
             logger.error(f"[KFC Replyer] LLM 调用失败: {raw_response}")
-            return LLMResponse.create_error_response(raw_response)
+            return False, "（回复生成失败）"
         
-        logger.debug(f"[KFC Replyer] LLM 响应 (model={model_name}):\n{raw_response}")
+        # 3. 清理并返回回复文本
+        reply_text = _clean_reply_text(raw_response)
         
-        # 3. 解析响应
+        logger.info(f"[KFC Replyer] 生成成功 (model={model_name}): {reply_text[:50]}...")
-        return _parse_response(raw_response)
+        
+        return True, reply_text
         
     except Exception as e:
         logger.error(f"[KFC Replyer] 生成失败: {e}")
         import traceback
         traceback.print_exc()
-        return LLMResponse.create_error_response(str(e))
+        return False, "（回复生成失败）"
 
 
-def _parse_response(raw_response: str) -> LLMResponse:
+def _clean_reply_text(raw_text: str) -> str:
-    """解析 LLM 响应"""
+    """
-    data = extract_and_parse_json(raw_response, strict=False)
+    清理回复文本
     
-    if not data or not isinstance(data, dict):
+    移除可能的前后缀、引号、markdown 标记等
-        logger.warning(f"[KFC Replyer] 无法解析 JSON: {raw_response[:200]}...")
+    """
-        return LLMResponse.create_error_response("无法解析响应格式")
+    text = raw_text.strip()
     
-    response = LLMResponse.from_dict(data)
+    # 移除可能的 markdown 代码块标记
+    if text.startswith("```") and text.endswith("```"):
+        lines = text.split("\n")
+        if len(lines) >= 3:
+            # 移除首尾的 ``` 行
+            text = "\n".join(lines[1:-1]).strip()
     
-    if response.thought:
+    # 移除首尾的引号（如果整个文本被引号包裹）
-        logger.info(
+    if (text.startswith('"') and text.endswith('"')) or \
-            f"[KFC Replyer] 解析成功: thought={response.thought[:50]}..., "
+       (text.startswith("'") and text.endswith("'")):
-            f"actions={[a.type for a in response.actions]}"
+        text = text[1:-1].strip()
-        )
-    else:
-        logger.warning("[KFC Replyer] 响应缺少 thought")
     
-    return response
+    # 移除可能的"你说："、"回复："等前缀
+    prefixes_to_remove = ["你说：", "你说:", "回复：", "回复:", "我说：", "我说:"]
+    for prefix in prefixes_to_remove:
+        if text.startswith(prefix):
+            text = text[len(prefix):].strip()
+            break
+    
+    return text