feat：精简表达选择，优化replyer表现，优化缓存命中率

prompts/zh-CN/maisaka_replyer.prompt

@@ -1,10 +1,7 @@
-在下面的内容中，标注 {bot_name}(你) 的发言是你自己的发言，请注意区分:
 {identity}
-{time_block}
-
 现在请你读读之前的聊天记录，把握当前的话题，然后给出日常且口语化的回复，
 {reply_style}
 你可以参考【回复信息参考】中的信息，但是视情况而定，不用完全遵守。
 {group_chat_attention_block}
 {replyer_at_block}
-请注意不要输出多余内容(包括不必要的前后缀，冒号，括号，表情包，普通 at 或 @等 )，只输出发言内容就好。
+请注意不要输出多余内容(包括不必要的前后缀，冒号，括号，表情包，@等 )，只输出发言内容就好。

pytests/test_context_message_fallback.py

@@ -1,4 +1,9 @@
-from src.common.data_models.message_component_data_model import ImageComponent, MessageSequence, TextComponent
+from src.common.data_models.message_component_data_model import (
+    ImageComponent,
+    MessageSequence,
+    ReplyComponent,
+    TextComponent,
+)
 from src.llm_models.payload_content.message import RoleType
 from src.maisaka.context_messages import _build_message_from_sequence
 from src.maisaka.message_adapter import build_visible_text_from_sequence
@@ -53,3 +58,16 @@ def test_visible_text_uses_image_placeholder_for_whitespace_content() -> None:
     )
 
     assert visible_text == "看这个[图片]"
+
+
+def test_visible_text_adds_body_marker_after_reply_component() -> None:
+    visible_text = build_visible_text_from_sequence(
+        MessageSequence(
+            [
+                ReplyComponent(target_message_id="75625487"),
+                TextComponent("你说是那就是"),
+            ]
+        )
+    )
+
+    assert visible_text == "[引用]quote_id=75625487\n[发言内容]你说是那就是"

pytests/test_maisaka_monitor_protocol.py

@@ -118,11 +118,16 @@ def test_legacy_replyer_builds_message_sequence_like_multimodal() -> None:
 
     assert len(request_messages) == 4
     assert request_messages[0].role.value == "system"
+    assert request_messages[0].get_text_content() == "legacy prompt"
     assert request_messages[1].role.value == "user"
     assert request_messages[1].get_text_content() == "[Alice]你好"
     assert request_messages[2].role.value == "user"
     assert request_messages[2].get_text_content() == "[Bob]在吗"
     assert request_messages[3].role.value == "user"
+    assert "当前时间：" in request_messages[3].get_text_content()
+    assert "【回复信息参考】" in request_messages[3].get_text_content()
+    assert "【最新推理】\n测试原因" in request_messages[3].get_text_content()
+    assert "请自然地回复。" in request_messages[3].get_text_content()
 
 
 @pytest.mark.asyncio

src/chat/replyer/maisaka_expression_selector.py

@@ -268,6 +268,11 @@ class MaisakaExpressionSelector:
         reply_reason: str,
         sub_agent_runner: Optional[SubAgentRunner],
     ) -> MaisakaExpressionSelectionResult:
+        del chat_history
+        del reply_message
+        del reply_reason
+        del sub_agent_runner
+
         if not session_id:
             logger.info("表达方式选择已跳过：缺少 session_id")
             return MaisakaExpressionSelectionResult()
@@ -280,48 +285,10 @@ class MaisakaExpressionSelector:
             logger.info(f"表达方式选择已跳过：本地候选不足，session_id={session_id}")
             return MaisakaExpressionSelectionResult()
 
-        if not global_config.expression.advanced_chosen:
         return self._build_direct_selection_result(
             session_id=session_id,
             candidates=candidates,
         )
 
-        if sub_agent_runner is None:
-            logger.info(f"表达方式选择已跳过：缺少 sub_agent_runner，session_id={session_id}")
-            return MaisakaExpressionSelectionResult()
-
-        logger.info(
-            f"表达方式选择开始：session_id={session_id} 候选数={len(candidates)} "
-            f"候选预览={self._format_candidate_preview(candidates)}"
-        )
-
-        selector_prompt = self._build_selector_prompt(
-            chat_history=chat_history,
-            reply_message=reply_message,
-            reply_reason=reply_reason,
-            candidates=candidates,
-        )
-        try:
-            raw_response = await sub_agent_runner(selector_prompt)
-        except Exception:
-            logger.exception("表达方式选择子代理执行失败")
-            return MaisakaExpressionSelectionResult()
-
-        selected_ids = self._parse_selected_ids(raw_response, candidates)
-        if not selected_ids:
-            logger.info(f"表达方式选择完成但未命中，session_id={session_id}")
-            return MaisakaExpressionSelectionResult()
-
-        selected_expressions = [candidate for candidate in candidates if candidate.get("id") in selected_ids]
-        self._update_last_active_time(selected_ids)
-        logger.info(
-            f"表达方式选择完成：session_id={session_id} 已选数={len(selected_ids)} "
-            f"selected_ids={selected_ids!r} 已选预览={self._format_candidate_preview(selected_expressions)}"
-        )
-        return MaisakaExpressionSelectionResult(
-            expression_habits=self._build_expression_habits_block(selected_expressions),
-            selected_expression_ids=selected_ids,
-        )
-
 
 maisaka_expression_selector = MaisakaExpressionSelector()

src/chat/replyer/maisaka_generator_base.py

@@ -18,6 +18,14 @@ from src.common.data_models.reply_generation_data_models import (
     ReplyGenerationResult,
     build_reply_monitor_detail,
 )
+from src.common.data_models.message_component_data_model import (
+    AtComponent,
+    EmojiComponent,
+    ImageComponent,
+    ReplyComponent,
+    TextComponent,
+    VoiceComponent,
+)
 from src.common.logger import get_logger
 from src.common.utils.utils_session import SessionUtils
 from src.config.config import global_config
@@ -100,6 +108,30 @@ class BaseMaisakaReplyGenerator:
         del message
         return ""
 
+    @staticmethod
+    def _strip_guided_reply_formatting(content: str) -> str:
+        """移除 guided_reply 可见文本中的引用包装，仅保留真正回复正文。"""
+
+        normalized_content = content.strip()
+        if not normalized_content:
+            return ""
+
+        reply_body_marker = "[发言内容]"
+        if normalized_content.startswith("[引用]quote_id=") or normalized_content.startswith("[引用消息]"):
+            marker_index = normalized_content.find(reply_body_marker)
+            if marker_index >= 0:
+                return normalized_content[marker_index + len(reply_body_marker) :].strip()
+
+            newline_index = normalized_content.find("\n")
+            if newline_index < 0:
+                return ""
+            normalized_content = normalized_content[newline_index + 1 :].lstrip()
+
+        if normalized_content.startswith(reply_body_marker):
+            normalized_content = normalized_content[len(reply_body_marker) :].lstrip()
+
+        return normalized_content.strip()
+
     def _extract_guided_bot_reply(self, message: SessionBackedMessage) -> str:
         # 只能根据结构化来源字段判断是否为 bot 自身写回的历史消息，
         # 不能依赖昵称/群名片等可控文本，避免误判和提示注入。
@@ -108,7 +140,9 @@ class BaseMaisakaReplyGenerator:
 
         plain_text = message.processed_plain_text.strip()
         _, body = parse_speaker_content(plain_text)
-        normalized_body = body.strip() or plain_text
+        normalized_body = self._strip_guided_reply_formatting(body) or self._strip_guided_reply_formatting(
+            plain_text
+        )
         return self._normalize_content(normalized_body) if normalized_body else ""
 
     def _build_target_message_block(self, reply_message: Optional[SessionMessage]) -> str:
@@ -118,18 +152,62 @@ class BaseMaisakaReplyGenerator:
         user_info = reply_message.message_info.user_info
         sender_name = user_info.user_cardname or user_info.user_nickname or user_info.user_id
         target_message_id = reply_message.message_id.strip() if reply_message.message_id else "未知"
-        target_content = self._normalize_content((reply_message.processed_plain_text or "").strip(), limit=300)
+        target_time = reply_message.timestamp.strftime("%Y-%m-%d %H:%M:%S")
+        target_content = self._normalize_content(self._build_target_message_content(reply_message), limit=300)
         if not target_content:
             target_content = "[无可见文本内容]"
 
         return (
             "【本次回复目标】\n"
-            f"- 目标消息ID：{target_message_id}\n"
-            f"- 发送者：{sender_name}\n"
-            f"- 消息内容：{target_content}\n"
-            "- 你这次要回复的就是这条目标消息，请结合整段上下文理解，但不要把其他历史消息当成当前回复对象。"
+            f"- msg_id：{target_message_id}\n"
+            f"- 时间：{target_time}\n"
+            f"- 用户名：{sender_name}\n"
+            f"- 发言内容：{target_content}\n\n"
+            "你这次要回复的就是这条目标消息，请结合整段上下文理解，但不要把其他历史消息当成当前回复对象。"
         )
 
+    @staticmethod
+    def _render_target_at_component(component: AtComponent) -> str:
+        target_name = component.target_user_cardname or component.target_user_nickname or component.target_user_id
+        return f"@{target_name}".strip()
+
+    def _build_target_message_content(self, reply_message: SessionMessage) -> str:
+        rendered_parts: List[str] = []
+
+        for component in reply_message.raw_message.components:
+            if isinstance(component, TextComponent):
+                if component.text:
+                    rendered_parts.append(component.text)
+                continue
+
+            if isinstance(component, ReplyComponent):
+                target_message_id = component.target_message_id.strip()
+                if target_message_id:
+                    rendered_parts.append(f"[引用:quote_id={target_message_id}]")
+                continue
+
+            if isinstance(component, AtComponent):
+                rendered_at = self._render_target_at_component(component)
+                if rendered_at:
+                    rendered_parts.append(rendered_at)
+                continue
+
+            if isinstance(component, ImageComponent):
+                rendered_parts.append(component.content.strip() or "[图片]")
+                continue
+
+            if isinstance(component, EmojiComponent):
+                rendered_parts.append(component.content.strip() or "[表情包]")
+                continue
+
+            if isinstance(component, VoiceComponent):
+                rendered_parts.append(component.content.strip() or "[语音消息]")
+
+        normalized_content = " ".join(part.strip() for part in rendered_parts if part and part.strip()).strip()
+        if normalized_content:
+            return normalized_content
+        return (reply_message.processed_plain_text or "").strip()
+
     @staticmethod
     def _get_chat_prompt_for_chat(chat_id: str, is_group_chat: Optional[bool]) -> str:
         """根据聊天流 ID 获取匹配的额外 prompt。"""
@@ -221,8 +299,10 @@ class BaseMaisakaReplyGenerator:
         expression_habits: str = "",
         stream_id: Optional[str] = None,
     ) -> str:
-        current_time = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
-        target_message_block = self._build_target_message_block(reply_message)
+        del reply_message
+        del reply_reason
+        del reference_info
+        del expression_habits
         session_id = self._resolve_session_id(stream_id)
 
         try:
@@ -231,16 +311,29 @@ class BaseMaisakaReplyGenerator:
                 bot_name=global_config.bot.nickname,
                 group_chat_attention_block=self._build_group_chat_attention_block(session_id),
                 replyer_at_block=self._build_replyer_at_block(),
-                time_block=f"当前时间：{current_time}",
                 identity=self._personality_prompt,
                 reply_style=global_config.personality.reply_style,
             )
         except Exception:
             system_prompt = "你是一个友好的 AI 助手，请根据聊天记录自然回复。"
 
-        sections: List[str] = []
+        return system_prompt
+
+    def _build_reply_instruction(self) -> str:
+        return "请自然地回复。不要输出多余说明、括号、@ 或额外标记，只输出实际要发送的内容。"
+
+    def _build_final_user_message(
+        self,
+        reply_message: Optional[SessionMessage],
+        reply_reason: str,
+        reference_info: str = "",
+        expression_habits: str = "",
+    ) -> str:
+        current_time = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+        sections: List[str] = [f"当前时间：{current_time}"]
         if expression_habits.strip():
             sections.append(expression_habits.strip())
+        target_message_block = self._build_target_message_block(reply_message)
         if target_message_block:
             sections.append(target_message_block)
         reply_reference_lines: List[str] = []
@@ -250,12 +343,8 @@ class BaseMaisakaReplyGenerator:
             reply_reference_lines.append(f"【参考信息】\n{reference_info.strip()}")
         if reply_reference_lines:
             sections.append("【回复信息参考】\n" + "\n\n".join(reply_reference_lines))
-        if not sections:
-            return system_prompt
-        return f"{system_prompt}\n\n" + "\n\n".join(sections)
-
-    def _build_reply_instruction(self) -> str:
-        return "请自然地回复。不要输出多余说明、括号、@ 或额外标记，只输出实际要发送的内容。"
+        sections.append(self._build_reply_instruction())
+        return "\n\n".join(sections)
 
     def _build_history_messages(
         self,
@@ -311,11 +400,16 @@ class BaseMaisakaReplyGenerator:
             expression_habits=expression_habits,
             stream_id=stream_id,
         )
-        instruction = self._build_reply_instruction()
+        final_user_message = self._build_final_user_message(
+            reply_message=reply_message,
+            reply_reason=reply_reason,
+            reference_info=reference_info,
+            expression_habits=expression_habits,
+        )
 
         messages.append(MessageBuilder().set_role(RoleType.System).add_text_content(system_prompt).build())
         messages.extend(self._build_history_messages(chat_history, enable_visual_message))
-        messages.append(MessageBuilder().set_role(RoleType.User).add_text_content(instruction).build())
+        messages.append(MessageBuilder().set_role(RoleType.User).add_text_content(final_user_message).build())
         return messages
 
     def _resolve_enable_visual_message(self, model_info: Optional[ModelInfo] = None) -> bool:
@@ -403,10 +497,10 @@ class BaseMaisakaReplyGenerator:
             result.error_message = "聊天历史为空"
             return finalize(False)
 
-        logger.info(
-            f"Maisaka 回复器开始生成: 流={stream_id} 原因={reply_reason!r} "
-            f"历史条数={len(chat_history)} 目标ID={reply_message.message_id if reply_message else None}"
-        )
+        # logger.info(
+        #     f"Maisaka 回复器开始生成: 流={stream_id} 原因={reply_reason!r} "
+        #     f"历史条数={len(chat_history)} 目标ID={reply_message.message_id if reply_message else None}"
+        # )
 
         filtered_history = [
             message
@@ -444,9 +538,9 @@ class BaseMaisakaReplyGenerator:
             else list(reply_context.selected_expression_ids)
         )
 
-        logger.info(
-            f"回复上下文完成 流={stream_id} 已选表达={result.selected_expression_ids!r}"
-        )
+        # logger.info(
+        #     f"回复上下文完成 流={stream_id} 已选表达={result.selected_expression_ids!r}"
+        # )
 
         prompt_started_at = time.perf_counter()
         try:

src/config/official_configs.py

@@ -720,15 +720,6 @@ class ExpressionConfig(ConfigBase):
     )
     """_wrap_表达学习配置列表，支持按聊天流配置"""
 
-    advanced_chosen: bool = Field(
-        default=False,
-        json_schema_extra={
-            "x-widget": "switch",
-            "x-icon": "sparkles",
-        },
-    )
-    """是否启用基于子代理的二次表达方式选择"""
-
     expression_groups: list[ExpressionGroup] = Field(
         default_factory=list,
         json_schema_extra={

src/maisaka/context_messages.py

@@ -90,7 +90,7 @@ def _append_reply_component(builder: MessageBuilder, component: ReplyComponent)
     if not target_message_id:
         return False
 
-    builder.add_text_content(f"[引用]quote_id={target_message_id}")
+    builder.add_text_content(f"[引用消息]{target_message_id}")
     return True
 
 
@@ -167,7 +167,7 @@ def _render_component_for_prompt(component: StandardMessageComponents) -> str:
         if target_content:
             return f"[回复消息: {target_content}]"
         target_message_id = component.target_message_id.strip()
-        return f"[引用]quote_id={target_message_id}" if target_message_id else "[回复消息]"
+        return f"[引用消息]{target_message_id}" if target_message_id else "[回复消息]"
 
     if isinstance(component, ForwardNodeComponent):
         return _build_forward_preview_block(component)

src/maisaka/history_post_processor.py

@@ -2,10 +2,9 @@
 
 from dataclasses import dataclass
 
-from .context_messages import AssistantMessage, LLMContextMessage, ToolResultMessage
+from .context_messages import AssistantMessage, LLMContextMessage
 from .history_utils import drop_leading_orphan_tool_results, drop_orphan_tool_results, normalize_tool_result_order
 
-TIMING_HISTORY_TOOL_NAMES = {"continue", "finish", "no_reply", "wait"}
 EARLY_TRIM_RATIO = 0.2
 
 
@@ -27,7 +26,6 @@ def process_chat_history_after_cycle(
     """在每轮结束后统一执行历史裁切与清理。"""
 
     processed_history = list(chat_history)
-    removed_timing_tool_count = _remove_early_timing_tool_records(processed_history)
     removed_assistant_thought_count = _remove_early_assistant_thoughts(processed_history)
 
     processed_history, orphan_removed_count = drop_orphan_tool_results(processed_history)
@@ -45,8 +43,7 @@ def process_chat_history_after_cycle(
     removed_overflow_count += leading_orphan_removed_count
     remaining_context_count = sum(1 for message in processed_history if message.count_in_context)
     removed_count = (
-        removed_timing_tool_count
-        + removed_assistant_thought_count
+        removed_assistant_thought_count
         + orphan_removed_count
         + removed_overflow_count
     )
@@ -59,41 +56,6 @@ def process_chat_history_after_cycle(
     )
 
 
-def _remove_early_timing_tool_records(chat_history: list[LLMContextMessage]) -> int:
-    """移除最早 20% 的门控/结束类工具链记录。"""
-
-    candidate_assistant_indexes = [
-        index
-        for index, message in enumerate(chat_history)
-        if _is_timing_tool_assistant_message(message)
-    ]
-    remove_count = int(len(candidate_assistant_indexes) * EARLY_TRIM_RATIO)
-    if remove_count <= 0:
-        return 0
-
-    removed_indexes = set(candidate_assistant_indexes[:remove_count])
-    removed_tool_call_ids = {
-        tool_call.call_id
-        for index in removed_indexes
-        for tool_call in chat_history[index].tool_calls
-        if tool_call.call_id
-    }
-
-    filtered_history: list[LLMContextMessage] = []
-    removed_total = 0
-    for index, message in enumerate(chat_history):
-        if index in removed_indexes:
-            removed_total += 1
-            continue
-        if isinstance(message, ToolResultMessage) and message.tool_call_id in removed_tool_call_ids:
-            removed_total += 1
-            continue
-        filtered_history.append(message)
-
-    chat_history[:] = filtered_history
-    return removed_total
-
-
 def _remove_early_assistant_thoughts(chat_history: list[LLMContextMessage]) -> int:
     """移除最早 20% 的非工具 assistant 思考内容。"""
 
@@ -122,8 +84,3 @@ def _remove_early_assistant_thoughts(chat_history: list[LLMContextMessage]) -> i
     return removed_total
 
 
-def _is_timing_tool_assistant_message(message: LLMContextMessage) -> bool:
-    if not isinstance(message, AssistantMessage) or not message.tool_calls:
-        return False
-
-    return all(tool_call.func_name in TIMING_HISTORY_TOOL_NAMES for tool_call in message.tool_calls)

src/maisaka/message_adapter.py

@@ -60,11 +60,23 @@ def build_visible_text_from_sequence(message_sequence: MessageSequence) -> str:
     """从消息片段序列提取可见文本。"""
 
     parts: list[str] = []
+    pending_reply_body_prefix = False
+
+    def append_visible_part(text: str) -> None:
+        nonlocal pending_reply_body_prefix
+        if not text:
+            return
+        if pending_reply_body_prefix:
+            parts.append(f"\n[发言内容]{text}")
+            pending_reply_body_prefix = False
+            return
+        parts.append(text)
+
     for component in message_sequence.components:
         if isinstance(component, TextComponent):
             match = SPEAKER_PREFIX_PATTERN.match(component.text or "")
             if not match:
-                parts.append(component.text)
+                append_visible_part(component.text)
                 continue
 
             normalized_parts: list[str] = []
@@ -75,24 +87,25 @@ def build_visible_text_from_sequence(message_sequence: MessageSequence) -> str:
                 normalized_parts.append(f"[msg_id:{message_id}]")
             normalized_parts.append(f"[{match.group('speaker')}]")
             normalized_parts.append(match.group("content"))
-            parts.append("".join(normalized_parts))
+            append_visible_part("".join(normalized_parts))
             continue
 
         if isinstance(component, EmojiComponent):
-            parts.append(component.content.strip() or "[表情包]")
+            append_visible_part(component.content.strip() or "[表情包]")
             continue
 
         if isinstance(component, ImageComponent):
-            parts.append(component.content.strip() or "[图片]")
+            append_visible_part(component.content.strip() or "[图片]")
             continue
 
         if isinstance(component, AtComponent):
-            parts.append(_render_at_component_text(component))
+            append_visible_part(_render_at_component_text(component))
             continue
 
         if isinstance(component, ReplyComponent):
             target_message_id = component.target_message_id.strip()
             if target_message_id:
-                parts.append(f"[引用]quote_id={target_message_id}")
+                parts.append(f"[引用消息]{target_message_id}")
+                pending_reply_body_prefix = True
 
     return "".join(parts)

src/maisaka/monitor_events.py

@@ -273,10 +273,14 @@ async def _broadcast(event: str, data: Dict[str, Any]) -> None:
             for connection in websocket_manager.connections.values()
             if subscription_key in connection.subscriptions
         )
-        logger.info(
-            f"[诊断] _broadcast: manager_id={id(websocket_manager)} "
-            f"总连接={total_connections} 订阅者={subscriber_count} event={event}"
-        )
+# The above code is using the Python logging module to log a diagnostic message. It is logging
+# information about the `_broadcast` function, including the `manager_id`, `total_connections`,
+# `subscriber_count`, and `event` variables. The `logger.info()` function is used to log the message
+# at the INFO level.
+        # logger.info(
+        #     f"[诊断] _broadcast: manager_id={id(websocket_manager)} "
+        #     f"总连接={total_connections} 订阅者={subscriber_count} event={event}"
+        # )
         await websocket_manager.broadcast_to_topic(
             domain=MONITOR_DOMAIN,
             topic=MONITOR_TOPIC,

src/maisaka/reasoning_engine.py

@@ -478,11 +478,11 @@ class MaisakaReasoningEngine:
                             )
                             planner_duration_ms = (time.time() - planner_started_at) * 1000
                             cycle_detail.time_records["planner"] = planner_duration_ms / 1000
-                            logger.info(
-                                f"{self._runtime.log_prefix} 规划器执行完成: "
-                                f"回合={round_index + 1} "
-                                f"耗时={cycle_detail.time_records['planner']:.3f} 秒"
-                            )
+                            # logger.info(
+                            #     f"{self._runtime.log_prefix} 规划器执行完成: "
+                            #     f"回合={round_index + 1} "
+                            #     f"耗时={cycle_detail.time_records['planner']:.3f} 秒"
+                            # )
                             reasoning_content = response.content or ""
                             if self._should_replace_reasoning(reasoning_content):
                                 response.content = "我应该根据我上面思考的内容进行反思，重新思考我下一步的行动，我需要分析当前场景，对话，以及我可以使用的工具，然后直接输出我的想法"
@@ -865,7 +865,7 @@ class MaisakaReasoningEngine:
             return False
 
         similarity = self._calculate_similarity(current_content, self._last_reasoning_content)
-        logger.info(f"{self._runtime.log_prefix} 思考内容相似度: {similarity:.2f}")
+        logger.debug(f"{self._runtime.log_prefix} 思考内容相似度: {similarity:.2f}")
         return similarity > 0.9
 
     @staticmethod