新增子 Agent 与技能系统管理：

- **子 Agent 支持**：
  - 新增 `dispatch_subagent` 工具，支持分配特定任务给子 Agent，由其独立操作。
  - 前端新增 `ChatSubAgent` 组件，支持子 Agent 任务状态展示（运行中/已完成）、工具调用进度及输出预览等。

- **技能管理扩展**：
  - 新增 Skill 编辑器页面（`SkillEditorView.vue`），支持创建、编辑、删除技能。
  - 完善技能工具权限分组选择功能，按类型灵活分配工具。

- **工具增强**：
  - 新增工具接口：`get_outline_detail`、`get_character_detail`、`get_chapter_detail`，用于获取详情内容。
  - 优化部分工具返回数据，仅输出基础信息以减少上下文体积。

- 后端添加 `SPECIAL_TOOLS` 支持，增强 `dispatch_subagent` 的特殊处理能力。

同时优化部分前后端交互，提升代码维护性与可读性。

backend/app/routers/chat.py

@@ -1,3 +1,4 @@
+import asyncio
 import json
 from datetime import datetime, timezone
 
@@ -12,7 +13,7 @@ from ..services.ai_provider import (
     stream_chat, simple_completion,
     build_assistant_message, build_tool_results_messages, ToolCall,
 )
-from ..services.tools import get_openai_tools, get_anthropic_tools, execute_tool, get_tools_description
+from ..services.tools import get_openai_tools, get_anthropic_tools, execute_tool, get_tools_description, SPECIAL_TOOLS
 
 router = APIRouter(prefix="/api/chat", tags=["chat"])
 
@@ -24,6 +25,7 @@ TOOL_LABELS: dict[str, str] = {
     "get_novel_info": "查看小说信息",
     "update_novel_info": "更新小说信息",
     "list_outlines": "查询大纲",
+    "get_outline_detail": "查看大纲详情",
     "create_outline": "创建大纲",
     "update_outline": "更新大纲",
     "delete_outline": "删除大纲",
@@ -31,17 +33,23 @@ TOOL_LABELS: dict[str, str] = {
     "get_world_setting": "查询世界观",
     "save_world_setting": "保存世界观",
     "list_characters": "查询角色",
+    "get_character_detail": "查看角色详情",
     "create_character": "创建角色",
     "update_character": "更新角色",
     "delete_character": "删除角色",
     "list_chapters": "查询章节",
+    "get_chapter_detail": "查看章节详情",
     "create_chapter": "创建章节",
     "update_chapter": "更新章节",
     "delete_chapter": "删除章节",
     "list_files": "查询文件",
     "read_file": "读取文件",
+    "dispatch_subagent": "派遣子Agent",
 }
 
+# 子 Agent 最大轮次
+MAX_SUBAGENT_ROUNDS = 5
+
 
 def _build_system_prompt(
     novel_id: int | None,
@@ -78,6 +86,21 @@ def _build_system_prompt(
         if not novel_id:
             parts.append("\n注意：工具操作需要关联到具体小说。如果用户需要使用工具，请提示他们先进入某本小说。")
 
+        # 注入可用 Skill 列表（用于 dispatch_subagent）
+        from sqlmodel import or_
+        skills = session.exec(
+            select(Skill).where(
+                or_(Skill.novel_id == novel_id, Skill.novel_id.is_(None))  # type: ignore
+            )
+        ).all()
+        if skills:
+            skill_lines = ["\n## 可用子Agent（Skill）"]
+            skill_lines.append("你可以使用 dispatch_subagent 工具派遣以下子Agent处理子任务：")
+            for sk in skills:
+                skill_lines.append(f"- skill_id={sk.id}  {sk.name}: {sk.description}")
+            skill_lines.append("当任务复杂时，考虑将子任务分派给专门的子Agent处理。")
+            parts.append("\n".join(skill_lines))
+
     return "\n".join(parts)
 
 
@@ -108,12 +131,342 @@ def _save_assistant_message(novel_id: int | None, full_response: str, tool_call_
             s.commit()
 
 
+async def _run_subagent(
+    queue: asyncio.Queue,
+    config: AIConfig,
+    skill: Skill,
+    task: str,
+    novel_id: int,
+    subagent_id: str = "",
+) -> str:
+    """运行子 Agent：独立 LLM 循环，通过 queue 发送事件。返回最终结果文本。
+    subagent_id 用于前端区分并行运行的多个子 Agent。"""
+    # 解析 Skill 的工具白名单
+    sub_allowed: list[str] | None = None
+    if skill.allowed_tools:
+        try:
+            parsed = json.loads(skill.allowed_tools)
+            if isinstance(parsed, list) and len(parsed) > 0:
+                sub_allowed = [t for t in parsed if t != "dispatch_subagent"]
+        except (json.JSONDecodeError, TypeError):
+            pass
+
+    # 构建子 Agent 系统提示词
+    sub_system_parts = [
+        f"你是子Agent「{skill.name}」。",
+        f"你的职责：{skill.description}" if skill.description else "",
+    ]
+    if skill.system_prompt:
+        sub_system_parts.append(skill.system_prompt)
+
+    with Session(engine) as s:
+        novel = s.get(Novel, novel_id)
+        if novel:
+            sub_system_parts.append(f"\n当前小说：《{novel.title}》（ID: {novel.id}）")
+        sub_tools_desc = get_tools_description(sub_allowed)
+        if sub_tools_desc:
+            sub_system_parts.append(f"\n## 你的工具能力\n{sub_tools_desc}")
+
+    sub_system = "\n".join(p for p in sub_system_parts if p)
+    sub_tools = _get_tools(config.provider, sub_allowed)
+    sub_messages: list[dict] = [{"role": "user", "content": task}]
+
+    full_response = ""
+    sid = subagent_id  # 简写
+
+    for _round in range(MAX_SUBAGENT_ROUNDS):
+        round_text = ""
+        round_tool_calls: list[ToolCall] = []
+
+        async for event in stream_chat(config, sub_messages, sub_system, sub_tools):
+            if event.text:
+                round_text += event.text
+                await queue.put({"subagent_chunk": event.text, "subagent_id": sid})
+            if event.tool_calls:
+                round_tool_calls = event.tool_calls
+
+        if not round_tool_calls:
+            full_response += round_text
+            break
+
+        full_response += round_text
+
+        assistant_msg = build_assistant_message(
+            config.provider, round_text, round_tool_calls
+        )
+        sub_messages.append(assistant_msg)
+
+        tool_results = []
+        for tc in round_tool_calls:
+            label = TOOL_LABELS.get(tc.name, tc.name)
+            await queue.put({"subagent_tool_call": {
+                "name": tc.name, "label": label, "arguments": tc.arguments,
+            }, "subagent_id": sid})
+            result = execute_tool(tc.name, tc.arguments, novel_id)
+            tool_results.append({"id": tc.id, "result": result})
+            await queue.put({"subagent_tool_result": {
+                "name": tc.name, "label": label, "result": result,
+            }, "subagent_id": sid})
+
+        result_msgs = build_tool_results_messages(config.provider, tool_results)
+        sub_messages.extend(result_msgs)
+
+    await queue.put({"subagent_done": full_response, "subagent_id": sid})
+    return full_response
+
+
+def _summarize_result(tool_name: str, result_json: str) -> str:
+    """将工具结果精简为简短摘要，减少上下文 token 占用。"""
+    try:
+        data = json.loads(result_json)
+    except (json.JSONDecodeError, TypeError):
+        return result_json[:100]
+
+    if "error" in data:
+        return f"错误: {data['error']}"
+
+    # 列表类工具 → 仅显示数量
+    if tool_name == "list_outlines":
+        return f"{data.get('total', 0)}条大纲"
+    if tool_name == "list_characters":
+        return f"{data.get('total', 0)}个角色"
+    if tool_name == "list_chapters":
+        return f"{data.get('total', 0)}个章节"
+    if tool_name == "list_files":
+        return f"{data.get('total', 0)}个文件"
+
+    # 创建类 → ID + 标题
+    if tool_name == "create_outline":
+        return f"已创建 #{data.get('id')} {data.get('summary', '')[:30]}"
+    if tool_name == "create_character":
+        return f"已创建 #{data.get('id')} {data.get('name', '')}"
+    if tool_name == "create_chapter":
+        return f"已创建 #{data.get('id')} {data.get('title', '')[:30]}"
+
+    # 更新类
+    if tool_name == "update_outline":
+        return f"已更新 #{data.get('id')}"
+    if tool_name == "update_character":
+        return f"已更新 #{data.get('id')} {data.get('name', '')}"
+    if tool_name == "update_chapter":
+        return f"已更新 #{data.get('id')}"
+    if tool_name == "update_novel_info":
+        return f"已更新小说信息"
+
+    # 删除类
+    if tool_name in ("delete_outline", "delete_character", "delete_chapter"):
+        return f"已删除 #{data.get('id')}"
+
+    # 排序
+    if tool_name == "reorder_outlines":
+        return f"已排序 {data.get('count', 0)}个节点"
+
+    # 世界观
+    if tool_name == "save_world_setting":
+        return f"已保存 ({data.get('content_length', 0)}字)"
+    if tool_name == "get_world_setting":
+        return f"{'有内容' if data.get('exists') else '暂无内容'}"
+
+    # 详情查看类
+    if tool_name == "get_outline_detail":
+        return f"大纲 #{data.get('id')} {data.get('summary', '')[:30]}"
+    if tool_name == "get_character_detail":
+        return f"角色 #{data.get('id')} {data.get('name', '')}"
+    if tool_name == "get_chapter_detail":
+        return f"章节 #{data.get('id')} {data.get('title', '')[:30]}"
+    if tool_name == "get_novel_info":
+        return f"小说: {data.get('title', '')}"
+
+    # 文件读取
+    if tool_name == "read_file":
+        return f"读取 {data.get('filename', '')} ({data.get('offset', 0)}-{data.get('offset', 0) + len(data.get('text', ''))}字/{data.get('total_length', 0)}字)"
+
+    # 其他 → 截取前 100 字符
+    return result_json[:100]
+
+
+async def _execute_tool_with_queue(
+    queue: asyncio.Queue,
+    tc: ToolCall,
+    config: AIConfig,
+    novel_id: int,
+) -> tuple[str, str]:
+    """执行单个工具（含子Agent），通过 queue 发事件。返回 (result, log_line)。
+    子 Agent 使用 tc.id 作为 subagent_id，用于前端区分并行的子 Agent。"""
+    label = TOOL_LABELS.get(tc.name, tc.name)
+
+    if tc.name == "dispatch_subagent":
+        skill_id = tc.arguments.get("skill_id")
+        task_desc = tc.arguments.get("task", "")
+        sub_skill = None
+        with Session(engine) as sub_s:
+            sub_skill = sub_s.get(Skill, skill_id)
+        if not sub_skill:
+            result = json.dumps({"error": f"Skill {skill_id} 不存在"}, ensure_ascii=False)
+        else:
+            await queue.put({"subagent_start": {
+                "skill_id": skill_id, "skill_name": sub_skill.name,
+                "task": task_desc, "subagent_id": tc.id,
+            }})
+            result = await _run_subagent(
+                queue, config, sub_skill, task_desc, novel_id, subagent_id=tc.id,
+            )
+        skill_name = sub_skill.name if sub_skill else "未知"
+        # 子 Agent 结果也精简（取前 200 字）
+        brief = result[:200] + "…" if len(result) > 200 else result
+        log_line = f"[子Agent: {skill_name}] 任务: {task_desc} → {brief}"
+    else:
+        result = execute_tool(tc.name, tc.arguments, novel_id)
+        await queue.put({"tool_result": {"name": tc.name, "label": label, "result": result}})
+        # 精简日志：仅记录摘要，避免完整 JSON 结果撑爆上下文
+        log_line = f"[工具调用: {label}] 参数: {json.dumps(tc.arguments, ensure_ascii=False)} → {_summarize_result(tc.name, result)}"
+
+    return result, log_line
+
+
+async def _execute_tools_parallel(
+    queue: asyncio.Queue,
+    tool_calls: list[ToolCall],
+    config: AIConfig,
+    novel_id: int,
+) -> tuple[list[dict], list[str]]:
+    """执行一组工具调用。子 Agent 并行执行，普通工具顺序执行。
+    返回 (tool_results, tool_call_logs)。"""
+    # 分离子 Agent 和普通工具
+    subagent_calls = [tc for tc in tool_calls if tc.name == "dispatch_subagent"]
+    normal_calls = [tc for tc in tool_calls if tc.name != "dispatch_subagent"]
+
+    results_map: dict[str, tuple[str, str]] = {}  # tc.id → (result, log_line)
+
+    # 普通工具顺序执行
+    for tc in normal_calls:
+        result, log_line = await _execute_tool_with_queue(queue, tc, config, novel_id)
+        results_map[tc.id] = (result, log_line)
+
+    # 子 Agent 并行执行
+    if subagent_calls:
+        async def _run_one(tc: ToolCall) -> tuple[str, tuple[str, str]]:
+            r, l = await _execute_tool_with_queue(queue, tc, config, novel_id)
+            return tc.id, (r, l)
+
+        tasks = [_run_one(tc) for tc in subagent_calls]
+        for coro in asyncio.as_completed(tasks):
+            tc_id, rl = await coro
+            results_map[tc_id] = rl
+
+    # 按原始顺序组装结果
+    tool_results = []
+    tool_call_logs = []
+    for tc in tool_calls:
+        result, log_line = results_map[tc.id]
+        tool_results.append({"id": tc.id, "result": result})
+        tool_call_logs.append(log_line)
+
+    return tool_results, tool_call_logs
+
+
+async def _chat_worker(
+    queue: asyncio.Queue,
+    config: AIConfig,
+    messages: list[dict],
+    system_prompt: str,
+    tools: list[dict] | None,
+    data: ChatRequest,
+):
+    """核心聊天逻辑 —— 运行在独立 Task 中，客户端断开也不会停止。"""
+    full_response = data.assistant_text or ""
+    usage_info: dict = {}
+    tool_call_logs: list[str] = []
+    pending_paused = False
+
+    try:
+        # ── 阶段一：执行待审批的工具调用 ──
+        if data.pending_tool_calls:
+            tool_calls = [
+                ToolCall(id=tc.id, name=tc.name, arguments=tc.arguments)
+                for tc in data.pending_tool_calls
+            ]
+            assistant_msg = build_assistant_message(
+                config.provider, data.assistant_text or "", tool_calls
+            )
+            messages.append(assistant_msg)
+
+            tool_results, logs = await _execute_tools_parallel(
+                queue, tool_calls, config, data.novel_id,
+            )
+            tool_call_logs.extend(logs)
+
+            result_msgs = build_tool_results_messages(config.provider, tool_results)
+            messages.extend(result_msgs)
+
+        # ── 阶段二：LLM 循环 ──
+        for _round in range(MAX_TOOL_ROUNDS):
+            round_text = ""
+            round_tool_calls: list[ToolCall] = []
+
+            async for event in stream_chat(config, messages, system_prompt, tools):
+                if event.text:
+                    round_text += event.text
+                    await queue.put({"content": event.text})
+                if event.usage:
+                    for k, v in event.usage.items():
+                        if v:
+                            usage_info[k] = usage_info.get(k, 0) + v
+                if event.tool_calls:
+                    round_tool_calls = event.tool_calls
+
+            if not round_tool_calls:
+                full_response += round_text
+                break
+
+            full_response += round_text
+            calls_data = [
+                {"id": tc.id, "name": tc.name,
+                 "label": TOOL_LABELS.get(tc.name, tc.name),
+                 "arguments": tc.arguments}
+                for tc in round_tool_calls
+            ]
+
+            if data.auto_approve_tools:
+                await queue.put({"tool_calls_auto": calls_data})
+
+                assistant_msg = build_assistant_message(
+                    config.provider, full_response, round_tool_calls
+                )
+                messages.append(assistant_msg)
+
+                tool_results, logs = await _execute_tools_parallel(
+                    queue, round_tool_calls, config, data.novel_id,
+                )
+                tool_call_logs.extend(logs)
+
+                result_msgs = build_tool_results_messages(config.provider, tool_results)
+                messages.extend(result_msgs)
+                continue
+            else:
+                # 审批模式：暂停
+                await queue.put({"tool_calls_pending": calls_data, "assistant_text": full_response})
+                await queue.put({"done": True, "pending": True, "usage": usage_info or None})
+                pending_paused = True
+                return  # 不保存到 DB
+
+    except Exception as e:
+        await queue.put({"error": str(e)})
+    finally:
+        if not pending_paused:
+            _save_assistant_message(data.novel_id, full_response, tool_call_logs)
+            await queue.put({"done": True, "usage": usage_info or None})
+        # 哨兵值：通知 SSE 生成器结束
+        await queue.put(None)
+
+
 @router.post("/stream")
 async def chat_stream(
     data: ChatRequest,
     session: Session = Depends(get_session),
 ):
-    """SSE 流式对话端点，支持 tool calling + 用户审批"""
+    """SSE 流式对话端点，支持 tool calling + 用户审批。
+    核心工作在独立 Task 中运行，客户端断开不会中断。"""
     config = session.exec(select(AIConfig)).first()
     if not config or not config.api_key:
         return StreamingResponse(
@@ -157,120 +510,23 @@ async def chat_stream(
     )
     tools = _get_tools(config.provider, allowed_tools) if use_tools else None
 
+    # 创建事件队列和后台工作任务
+    queue: asyncio.Queue = asyncio.Queue()
+    asyncio.create_task(
+        _chat_worker(queue, config, messages, system_prompt, tools, data)
+    )
+
     async def generate():
-        nonlocal messages
-
-        full_response = data.assistant_text or ""
-        usage_info: dict = {}
-        # 记录本次对话中的工具调用摘要（用于保存到 DB）
-        tool_call_logs: list[str] = []
-
+        """SSE 生成器：从队列读取事件。客户端断开时后台任务继续运行。"""
         try:
-            # ── 阶段一：如果有待审批的工具调用，先执行它们 ──
-            if data.pending_tool_calls:
-                tool_calls = [
-                    ToolCall(id=tc.id, name=tc.name, arguments=tc.arguments)
-                    for tc in data.pending_tool_calls
-                ]
-
-                # 构建 assistant 消息（含工具调用）
-                assistant_msg = build_assistant_message(
-                    config.provider, data.assistant_text or "", tool_calls
-                )
-                messages.append(assistant_msg)
-
-                # 执行每个工具，发送结果
-                tool_results = []
-                for tc in tool_calls:
-                    label = TOOL_LABELS.get(tc.name, tc.name)
-                    result = execute_tool(tc.name, tc.arguments, data.novel_id)
-                    tool_results.append({"id": tc.id, "result": result})
-                    tool_call_logs.append(f"[工具调用: {label}] 参数: {json.dumps(tc.arguments, ensure_ascii=False)} → 结果: {result}")
-                    yield _sse({"tool_result": {
-                        "name": tc.name,
-                        "label": label,
-                        "result": result,
-                    }})
-
-                # 追加工具结果消息
-                result_msgs = build_tool_results_messages(config.provider, tool_results)
-                messages.extend(result_msgs)
-
-            # ── 阶段二：LLM 循环 ──
-            for _round in range(MAX_TOOL_ROUNDS):
-                round_text = ""
-                round_tool_calls = []
-
-                async for event in stream_chat(config, messages, system_prompt, tools):
-                    if event.text:
-                        round_text += event.text
-                        yield _sse({"content": event.text})
-                    if event.usage:
-                        for k, v in event.usage.items():
-                            if v:
-                                usage_info[k] = usage_info.get(k, 0) + v
-                    if event.tool_calls:
-                        round_tool_calls = event.tool_calls
-
-                # 没有工具调用 → 完成
-                if not round_tool_calls:
-                    full_response += round_text
+            while True:
+                event = await queue.get()
+                if event is None:
                     break
-
-                # 有工具调用
-                full_response += round_text
-                calls_data = [
-                    {
-                        "id": tc.id,
-                        "name": tc.name,
-                        "label": TOOL_LABELS.get(tc.name, tc.name),
-                        "arguments": tc.arguments,
-                    }
-                    for tc in round_tool_calls
-                ]
-
-                if data.auto_approve_tools:
-                    # 自动执行模式：直接执行工具，不暂停
-                    yield _sse({"tool_calls_auto": calls_data})
-
-                    assistant_msg = build_assistant_message(
-                        config.provider, full_response, round_tool_calls
-                    )
-                    messages.append(assistant_msg)
-
-                    tool_results = []
-                    for tc in round_tool_calls:
-                        label = TOOL_LABELS.get(tc.name, tc.name)
-                        result = execute_tool(tc.name, tc.arguments, data.novel_id)
-                        tool_results.append({"id": tc.id, "result": result})
-                        tool_call_logs.append(f"[工具调用: {label}] 参数: {json.dumps(tc.arguments, ensure_ascii=False)} → 结果: {result}")
-                        yield _sse({"tool_result": {
-                            "name": tc.name,
-                            "label": label,
-                            "result": result,
-                        }})
-
-                    result_msgs = build_tool_results_messages(config.provider, tool_results)
-                    messages.extend(result_msgs)
-                    # 继续循环，让 LLM 基于工具结果生成回复
-                    continue
-                else:
-                    # 审批模式：暂停，发给前端审批
-                    yield _sse({
-                        "tool_calls_pending": calls_data,
-                        "assistant_text": full_response,
-                    })
-                    yield _sse({"done": True, "pending": True, "usage": usage_info or None})
-                    return  # 暂停，不保存到 DB
-
-        except Exception as e:
-            yield _sse({"error": str(e)})
-        finally:
-            # 无论连接是否断开，都保存已有的回复和工具调用到 DB
-            # 这确保刷新页面后 loadHistory 能恢复正确状态
-            _save_assistant_message(data.novel_id, full_response, tool_call_logs)
-
-        yield _sse({"done": True, "usage": usage_info or None})
+                yield _sse(event)
+        except (asyncio.CancelledError, GeneratorExit):
+            # 客户端断开连接 — 后台任务继续运行，最终会保存到 DB
+            pass
 
     return StreamingResponse(
         generate(),
@@ -354,18 +610,16 @@ async def compress_context(
     if len(all_msgs) <= KEEP_RECENT_ROUNDS * 2:
         return {"compressed": False, "reason": "消息太少，无需压缩"}
 
-    # 分割：旧消息（需要压缩）+ 最近消息（保留原文）
+    # 分割：旧消息（将被删除）+ 最近消息（保留原文）
     keep_count = KEEP_RECENT_ROUNDS * 2
     old_msgs = all_msgs[:-keep_count]
     recent_msgs = all_msgs[-keep_count:]
 
-    # 检查第一条旧消息是否已是摘要（避免重复压缩无效内容）
-    # 构建需要总结的对话文本
+    # 将 **所有消息**（包括最近保留的）都给 LLM 看，确保摘要完整准确
     conversation_text = ""
-    for msg in old_msgs:
+    for msg in all_msgs:
         role_label = "用户" if msg.role == "user" else "AI助手"
         content = msg.content
-        # 如果是已压缩的摘要，标注出来
         if content.startswith("[对话摘要]"):
             conversation_text += f"[之前的摘要]{content[5:]}\n\n"
         else:
@@ -373,10 +627,11 @@ async def compress_context(
 
     # 调用 LLM 生成摘要
     summary_prompt = (
-        "你是一个对话摘要助手。请将以下对话历史压缩为简洁的摘要，保留关键信息：\n"
+        "你是一个对话摘要助手。请将以下完整对话历史压缩为简洁的摘要，保留关键信息：\n"
         "1. 用户提出的核心需求和决策\n"
         "2. AI 执行的重要操作及结果（如创建/修改了哪些大纲、角色、章节等）\n"
-        "3. 达成的共识和待办事项\n\n"
+        "3. 达成的共识和待办事项\n"
+        "4. 当前工作进展和下一步计划\n\n"
         "只输出摘要内容，不要加前缀或解释。用简洁的条目列表形式。"
     )
     summary_messages = [
@@ -392,12 +647,14 @@ async def compress_context(
     for msg in old_msgs:
         session.delete(msg)
 
-    # 插入摘要消息（时间设为最早保留消息之前）
+    # 插入摘要消息，时间设为保留消息之前（确保排在最前面）
+    from datetime import timedelta
+    earliest_kept = recent_msgs[0].created_at
     summary_msg = ChatMessage(
         novel_id=novel_id,
         role="assistant",
         content=f"[对话摘要]\n{summary}",
-        created_at=recent_msgs[0].created_at,
+        created_at=earliest_kept - timedelta(seconds=1),
     )
     session.add(summary_msg)
     session.commit()