feat：根据人格生成言语风格

src/main.py

@@ -108,6 +108,9 @@ class MainSystem:
         )
         logger.success("个体特征初始化成功")
 
+        # 初始化表达方式
+        await expression_learner.extract_and_store_personality_expressions()
+
         try:
             # 启动全局消息管理器 (负责消息发送/排队)
             await message_manager.start()

src/plugins/heartFC_chat/expressors/exprssion_learner.py

@@ -33,10 +33,32 @@ def init_prompt() -> None:
 当"想说明某个观点，但懒得明说"，使用"懂的都懂"
 当"想搞笑的表现高深的感觉"，使用"文言文句式"
 
+注意不要总结你自己的发言
 现在请你概括
 """
     Prompt(learn_expression_prompt, "learn_expression_prompt")
 
+    personality_expression_prompt = """
+{personality}
+
+请从以上人设中总结出这个角色可能的语言风格
+思考回复语法，长度和情感
+思考有没有特殊的梗，一并总结成语言风格
+总结成如下格式的规律，总结的内容要详细，但具有概括性：
+当"xxx"时，可以"xxx", xxx不超过10个字
+
+例如：
+当"表示十分惊叹"时，使用"我嘞个xxxx"
+当"表示讽刺的赞同，不想讲道理"时，使用"对对对"
+当"想表达某个观点，但不想明说"，使用"反讽的句式"
+当"想说明某个观点，但懒得明说"，使用"懂的都懂"
+当"想搞笑的表现高深的感觉"，使用"文言文句式"
+
+现在请你概括
+"""
+    Prompt(personality_expression_prompt, "personality_expression_prompt")
+
+peronality = "情绪敏感，有时候有些搞怪幽默, 是一个女大学生，现在在读大二，你会刷贴吧"
 
 class ExpressionLearner:
     def __init__(self) -> None:
@@ -47,14 +69,22 @@ class ExpressionLearner:
             request_type="response_heartflow",
         )
 
-    async def get_expression_by_chat_id(self, chat_id: str) -> List[Dict[str, str]]:
+    async def get_expression_by_chat_id(self, chat_id: str) -> Tuple[List[Dict[str, str]], List[Dict[str, str]]]:
-        """从/data/expression/对应chat_id/expressions.json中读取表达方式"""
+        """
-        file_path: str = os.path.join("data", "expression", str(chat_id), "expressions.json")
+        读取/data/expression/learnt/{chat_id}/expressions.json和/data/expression/personality/expressions.json
-        if not os.path.exists(file_path):
+        返回(learnt_expressions, personality_expressions)
-            return []
+        """
-        with open(file_path, "r", encoding="utf-8") as f:
+        learnt_file = os.path.join("data", "expression", "learnt", str(chat_id), "expressions.json")
-            expressions: List[dict] = json.load(f)
+        personality_file = os.path.join("data", "expression", "personality", "expressions.json")
-        return expressions
+        learnt_expressions = []
+        personality_expressions = []
+        if os.path.exists(learnt_file):
+            with open(learnt_file, "r", encoding="utf-8") as f:
+                learnt_expressions = json.load(f)
+        if os.path.exists(personality_file):
+            with open(personality_file, "r", encoding="utf-8") as f:
+                personality_expressions = json.load(f)
+        return learnt_expressions, personality_expressions
 
     def is_similar(self, s1: str, s2: str) -> bool:
         """
@@ -85,7 +115,7 @@ class ExpressionLearner:
             chat_dict[chat_id].append({"situation": situation, "style": style})
         # 存储到/data/expression/对应chat_id/expressions.json
         for chat_id, expr_list in chat_dict.items():
-            dir_path = os.path.join("data", "expression", str(chat_id))
+            dir_path = os.path.join("data", "expression", "learnt", str(chat_id))
             os.makedirs(dir_path, exist_ok=True)
             file_path = os.path.join(dir_path, "expressions.json")
             # 若已存在，先读出合并
@@ -188,6 +218,38 @@ class ExpressionLearner:
             expressions.append((chat_id, situation, style))
         return expressions
 
+    async def extract_and_store_personality_expressions(self):
+        """
+        检查data/expression/personality目录，不存在则创建。
+        用peronality变量作为chat_str，调用LLM生成表达风格，解析后count=100，存储到expressions.json。
+        """
+        dir_path = os.path.join("data", "expression", "personality")
+        os.makedirs(dir_path, exist_ok=True)
+        file_path = os.path.join(dir_path, "expressions.json")
+
+        # 构建prompt
+        prompt = await global_prompt_manager.format_prompt(
+            "personality_expression_prompt",
+            personality=peronality,
+        )
+        logger.info(f"个性表达方式提取prompt: {prompt}")
+        response, _ = await self.express_learn_model.generate_response_async(prompt)
+        logger.info(f"个性表达方式提取response: {response}")
+        # chat_id用personality
+        expressions = self.parse_expression_response(response, "personality")
+        # 转为dict并count=100
+        result = []
+        for _, situation, style in expressions:
+            result.append({"situation": situation, "style": style, "count": 100})
+        # 超过50条时随机删除多余的，只保留50条
+        if len(result) > 50:
+            remove_count = len(result) - 50
+            remove_indices = set(random.sample(range(len(result)), remove_count))
+            result = [item for idx, item in enumerate(result) if idx not in remove_indices]
+        with open(file_path, "w", encoding="utf-8") as f:
+            json.dump(result, f, ensure_ascii=False, indent=2)
+        logger.info(f"已写入{len(result)}条表达到{file_path}")
+
 
 init_prompt()
 

src/plugins/heartFC_chat/heartflow_prompt_builder.py

@@ -246,14 +246,22 @@ async def _build_prompt_focus(
         structured_info_prompt = ""
 
     # 从/data/expression/对应chat_id/expressions.json中读取表达方式
-    expressions = await expression_learner.get_expression_by_chat_id(chat_stream.stream_id)
+    learnt_expressions, personality_expressions = await expression_learner.get_expression_by_chat_id(chat_stream.stream_id)
-    if expressions:
+    language_habits = []
-        language_habits = []
+    # 1. learnt_expressions加权随机选5条
-        for expr in expressions:
+    if learnt_expressions:
-            # print(f"expr: {expr}")
+        weights = [expr["count"] for expr in learnt_expressions]
+        selected_learnt = weighted_sample_no_replacement(learnt_expressions, weights, 5)
+        for expr in selected_learnt:
             if isinstance(expr, dict) and "situation" in expr and "style" in expr:
                 language_habits.append(f"当{expr['situation']}时，使用 {expr['style']}")
-    else:
+    # 2. personality_expressions随机选1条
+    if personality_expressions:
+        expr = random.choice(personality_expressions)
+        if isinstance(expr, dict) and "situation" in expr and "style" in expr:
+            language_habits.append(f"当{expr['situation']}时，使用 {expr['style']}")
+    # 3. 如果都没有，给默认
+    if not language_habits:
         language_habits = [
             "当表示惊叹时，使用 我嘞个xxxx",
             "当表示惊讶无语是，使用 不是？",
@@ -264,8 +272,8 @@ async def _build_prompt_focus(
             "当你想表达某个观点，但不想明说，可以使用反讽",
             "当你想说明某个观点，但懒得明说，使用 懂的都懂",
         ]
-    selected = random.sample(language_habits, 8)
+        language_habits = random.sample(language_habits, 6)
-    language_habits_str = "\n".join(selected)
+    language_habits_str = "\n".join(language_habits)
 
     logger.debug("开始构建 focus prompt")
 
@@ -855,5 +863,37 @@ class PromptBuilder:
             return "[构建 Planner Prompt 时出错]"
 
 
+def weighted_sample_no_replacement(items, weights, k):
+    """
+    加权且不放回地随机抽取k个元素。
+
+    参数：
+        items: 待抽取的元素列表
+        weights: 每个元素对应的权重（与items等长，且为正数）
+        k: 需要抽取的元素个数
+    返回：
+        selected: 按权重加权且不重复抽取的k个元素组成的列表
+
+    实现思路：
+        每次从当前池中按权重加权随机选出一个元素，选中后将其从池中移除，重复k次。
+        这样保证了：
+        1. count越大被选中概率越高
+        2. 不会重复选中同一个元素
+    """
+    selected = []
+    pool = list(zip(items, weights))
+    for _ in range(min(k, len(pool))):
+        total = sum(w for _, w in pool)
+        r = random.uniform(0, total)
+        upto = 0
+        for idx, (item, weight) in enumerate(pool):
+            upto += weight
+            if upto >= r:
+                selected.append(item)
+                pool.pop(idx)
+                break
+    return selected
+
+
 init_prompt()
 prompt_builder = PromptBuilder()

src/plugins/utils/chat_message_builder.py

@@ -248,7 +248,13 @@ async def _build_readable_messages_internal(
             new_content += content[last_end:]
             content = new_content
         
-        message_details_raw.append((timestamp, person_name, content))
+        target_str = "这是QQ的一个功能，用于提及某人，但没那么明显"
+        if target_str in content:
+            if random.random() < 0.6:
+                content = content.replace(target_str, "")
+        
+        if content != "":
+            message_details_raw.append((timestamp, person_name, content))
 
     if not message_details_raw:
         return "", []