feat：进一步优化缓存

src/chat/replyer/maisaka_generator_base.py

@@ -108,30 +108,6 @@ 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。
@@ -140,9 +116,7 @@ class BaseMaisakaReplyGenerator:
 
         plain_text = message.processed_plain_text.strip()
         _, body = parse_speaker_content(plain_text)
-        normalized_body = self._strip_guided_reply_formatting(body) or self._strip_guided_reply_formatting(
+        normalized_body = body.strip()
-            plain_text
-        )
         return self._normalize_content(normalized_body) if normalized_body else ""
 
     def _build_target_message_block(self, reply_message: Optional[SessionMessage]) -> str:

src/maisaka/builtin_tool/context.py

@@ -244,7 +244,10 @@ class BuiltinToolRuntimeContext:
         history_message = SessionBackedMessage.from_session_message(
             message,
             raw_message=build_prefixed_message_sequence(message.raw_message, planner_prefix),
-            visible_text=build_session_message_visible_text(message),
+            visible_text=build_session_message_visible_text(
+                message,
+                include_reply_components=source_kind != "guided_reply",
+            ),
             source_kind=source_kind,
         )
         self.runtime._chat_history.append(history_message)

src/maisaka/chat_loop_service.py

@@ -650,7 +650,6 @@ class MaisakaChatLoopService:
 
         selected_indices.reverse()
         selected_history = [filtered_history[index] for index in selected_indices]
-        selected_history, _ = MaisakaChatLoopService._hide_early_assistant_messages(selected_history)
         selected_history, _ = drop_orphan_tool_results(selected_history)
         selected_history, _ = normalize_tool_result_order(selected_history)
         tool_message_count = sum(1 for message in selected_history if isinstance(message, ToolResultMessage))
@@ -709,38 +708,3 @@ class MaisakaChatLoopService:
             return resolve_enable_visual_planner()
         return True
 
-    @staticmethod
-    def _hide_early_assistant_messages(
-        selected_history: List[LLMContextMessage],
-    ) -> tuple[List[LLMContextMessage], int]:
-        """隐藏上下文中最早 50% 的 assistant 文本消息，但保留工具调用链路。"""
-
-        assistant_indices = [
-            index
-            for index, message in enumerate(selected_history)
-            if isinstance(message, AssistantMessage)
-        ]
-        hidden_assistant_count = len(assistant_indices) // 2
-        if hidden_assistant_count <= 0:
-            return selected_history, 0
-
-        removed_assistant_indices = set(assistant_indices[:hidden_assistant_count])
-
-        filtered_history: List[LLMContextMessage] = []
-        for index, message in enumerate(selected_history):
-            if index in removed_assistant_indices:
-                if not message.tool_calls:
-                    continue
-                filtered_history.append(
-                    AssistantMessage(
-                        content="",
-                        timestamp=message.timestamp,
-                        tool_calls=list(message.tool_calls),
-                        source_kind=message.source_kind,
-                    )
-                )
-                continue
-            filtered_history.append(message)
-
-        return filtered_history, hidden_assistant_count
-

src/maisaka/history_post_processor.py

@@ -1,11 +1,13 @@
 """Maisaka 历史消息轮次结束后处理。"""
 
 from dataclasses import dataclass
+from math import ceil
 
 from .context_messages import AssistantMessage, LLMContextMessage
 from .history_utils import drop_leading_orphan_tool_results, drop_orphan_tool_results, normalize_tool_result_order
 
-EARLY_TRIM_RATIO = 0.2
+EARLY_TRIM_RATIO = 0.3
+TRIM_THRESHOLD_RATIO = 1.2
 
 
 @dataclass(slots=True)
@@ -26,27 +28,32 @@ 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_assistant_thought_count = _remove_early_assistant_thoughts(processed_history)
+    processed_history, normalized_removed_count, moved_tool_result_count = _normalize_history_structure(
-
+        processed_history
-    processed_history, orphan_removed_count = drop_orphan_tool_results(processed_history)
-    processed_history, moved_tool_result_count = normalize_tool_result_order(processed_history)
-    remaining_context_count = sum(1 for message in processed_history if message.count_in_context)
-    removed_overflow_count = 0
-
-    while remaining_context_count > max_context_size and processed_history:
-        removed_message = processed_history.pop(0)
-        removed_overflow_count += 1
-        if removed_message.count_in_context:
-            remaining_context_count -= 1
-
-    processed_history, leading_orphan_removed_count = drop_leading_orphan_tool_results(processed_history)
-    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_assistant_thought_count
-        + orphan_removed_count
-        + removed_overflow_count
     )
+    remaining_context_count = sum(1 for message in processed_history if message.count_in_context)
+
+    compact_removed_count = 0
+    trim_threshold = ceil(max_context_size * TRIM_THRESHOLD_RATIO)
+    if remaining_context_count > trim_threshold:
+        removed_early_message_count = _remove_early_history_messages(processed_history)
+        processed_history, removed_after_message_trim_count, moved_after_message_trim_count = (
+            _normalize_history_structure(processed_history)
+        )
+        removed_assistant_thought_count = _remove_early_assistant_thoughts(processed_history)
+        processed_history, removed_after_thought_trim_count, moved_after_thought_trim_count = (
+            _normalize_history_structure(processed_history)
+        )
+        compact_removed_count = (
+            removed_early_message_count
+            + removed_after_message_trim_count
+            + removed_assistant_thought_count
+            + removed_after_thought_trim_count
+        )
+        moved_tool_result_count += moved_after_message_trim_count + moved_after_thought_trim_count
+
+    remaining_context_count = sum(1 for message in processed_history if message.count_in_context)
+    removed_count = normalized_removed_count + compact_removed_count
     changed_count = removed_count + moved_tool_result_count
     return HistoryPostProcessResult(
         history=processed_history,
@@ -56,8 +63,34 @@ def process_chat_history_after_cycle(
     )
 
 
+def _normalize_history_structure(
+    chat_history: list[LLMContextMessage],
+) -> tuple[list[LLMContextMessage], int, int]:
+    """规范化历史消息结构，保证工具调用链符合 LLM 消息协议。"""
+
+    processed_history, orphan_removed_count = drop_orphan_tool_results(chat_history)
+    processed_history, moved_tool_result_count = normalize_tool_result_order(processed_history)
+    processed_history, leading_orphan_removed_count = drop_leading_orphan_tool_results(processed_history)
+    return (
+        processed_history,
+        orphan_removed_count + leading_orphan_removed_count,
+        moved_tool_result_count,
+    )
+
+
+def _remove_early_history_messages(chat_history: list[LLMContextMessage]) -> int:
+    """移除最早 30% 的全部历史消息。"""
+
+    remove_count = int(len(chat_history) * EARLY_TRIM_RATIO)
+    if remove_count <= 0:
+        return 0
+
+    del chat_history[:remove_count]
+    return remove_count
+
+
 def _remove_early_assistant_thoughts(chat_history: list[LLMContextMessage]) -> int:
-    """移除最早 20% 的非工具 assistant 思考内容。"""
+    """移除最早 30% 的非工具 assistant 思考内容。"""
 
     candidate_indexes = [
         index

src/maisaka/history_utils.py

@@ -2,7 +2,7 @@
 
 from typing import TYPE_CHECKING
 
-from src.common.data_models.message_component_data_model import MessageSequence, TextComponent
+from src.common.data_models.message_component_data_model import MessageSequence, ReplyComponent, TextComponent
 
 from .context_messages import AssistantMessage, LLMContextMessage, ToolResultMessage
 from .message_adapter import build_visible_text_from_sequence, clone_message_sequence, format_speaker_content
@@ -28,6 +28,8 @@ def build_prefixed_message_sequence(
 def build_session_message_visible_text(
     message: "SessionMessage",
     source_sequence: MessageSequence | None = None,
+    *,
+    include_reply_components: bool = True,
 ) -> str:
     """将真实会话消息转换为 Maisaka 可见文本。"""
 
@@ -46,6 +48,8 @@ def build_session_message_visible_text(
         )
     )
     for component in clone_message_sequence(normalized_sequence).components:
+        if not include_reply_components and isinstance(component, ReplyComponent):
+            continue
         visible_sequence.components.append(component)
     return build_visible_text_from_sequence(visible_sequence).strip()
 

src/maisaka/runtime.py

@@ -233,7 +233,10 @@ class MaisakaHeartFlowChatting:
             history_message = SessionBackedMessage.from_session_message(
                 message,
                 raw_message=build_prefixed_message_sequence(message.raw_message, planner_prefix),
-                visible_text=build_session_message_visible_text(message),
+                visible_text=build_session_message_visible_text(
+                    message,
+                    include_reply_components=source_kind != "guided_reply",
+                ),
                 source_kind=source_kind,
             )
             self._chat_history.append(history_message)