feat: v1.1.0 — 上下文引擎重构 + 技能系统升级 + SOUL.md 支持

- 版本号更新: 1.0.0 → 1.1.0，全量同步 package.json/lock/README/docs/UI - 上下文长度自动检测: 切换模型时从 model_info 读取实际 context_length - SOUL.md 支持: 每次发送自动扫描工作空间 SOUL.md，注入为首条 system 消息且不可压缩 - 系统提示词卡片: AI 回复顶部可折叠展示实际发送给模型的完整 system prompt - Token 校准: 利用 Ollama 返回的实际计数动态修正估算器(EMA) - 消息重要性评分: 纯规则打分，trim/summarize 按重要性而非时间顺序 - 结构化压缩: LLM 压缩输出 JSON 格式(topics/decisions/pendingTasks/constraints) - 技能系统 v1.1: 语义匹配(embedding) + 参数自优化 + 未使用衰减 + 技能链合并 - 修复: 100+ TypeScript strict 模式编译错误清零
2026-06-04 21:47:18 +08:00
parent b12837433b
commit db6cb8d2cc
27 changed files with 646 additions and 139 deletions
@@ -0,0 +1 @@
+registry=https://registry.npmmirror.com
@@ -14,7 +14,7 @@
 </p>

 <p align="center">
-  <img src="https://img.shields.io/badge/version-v1--stable--release-E8734A?style=flat-square" alt="version">
+  <img src="https://img.shields.io/badge/version-v1.1.0-E8734A?style=flat-square" alt="version">
  <img src="https://img.shields.io/badge/electron-33+-47848F?style=flat-square&logo=electron" alt="electron">
  <img src="https://img.shields.io/badge/typescript-5.7+-3178C6?style=flat-square&logo=typescript" alt="typescript">
  <img src="https://img.shields.io/badge/license-MIT-green?style=flat-square" alt="license">
@@ -212,7 +212,7 @@ SQLite (sql.js WASM)，7 张表，WAL 模式 + FTS5 全文搜索：
 ```bash
 git clone https://gitee.com/thzxx/metona-ollama-desktop.git
 cd metona-ollama-desktop
-git checkout desktop-v1-stable-release
+git checkout desktop-v1.1.0
 npm config set registry https://registry.npmmirror.com
 npm install
 npm start
@@ -225,7 +225,7 @@ npm start
 ELECTRON_MIRROR=https://npmmirror.com/mirrors/electron/ npm run dist
 ```

-产出：`release/Metona Ollama Setup v1-stable-release.exe`
+产出：`release/Metona Ollama Setup v1.1.0.exe`

 ## 🛠️ 常用命令

@@ -433,7 +433,7 @@ SQLite (sql.js WASM), 7 tables, WAL mode + FTS5 full-text search:
 ```bash
 git clone https://gitee.com/thzxx/metona-ollama-desktop.git
 cd metona-ollama-desktop
-git checkout desktop-v1-stable-release
+git checkout desktop-v1.1.0
 npm config set registry https://registry.npmmirror.com
 npm install
 npm start
@@ -446,7 +446,7 @@ npm start
 ELECTRON_MIRROR=https://npmmirror.com/mirrors/electron/ npm run dist
 ```

-Output: `release/Metona Ollama Setup v1-stable-release.exe`
+Output: `release/Metona Ollama Setup v1.1.0.exe`

 ## 🛠️ Common Commands

@@ -14,7 +14,7 @@
 # 1. 克隆并安装依赖
 git clone https://gitee.com/thzxx/metona-ollama-desktop.git
 cd metona-ollama-desktop
-git checkout desktop-v1-stable-release
+git checkout desktop-v1.1.0
 npm config set registry https://registry.npmmirror.com
 npm install

@@ -92,7 +92,7 @@ npm run dist           # NSIS 安装包

 | 文件 | 说明 |
 |------|------|
-| `release/Metona Ollama Setup v1-stable-release.exe` | NSIS 安装包（可选目录、创建快捷方式） |
+| `release/Metona Ollama Setup v1.1.0.exe` | NSIS 安装包（可选目录、创建快捷方式） |

 > v3.2.0 起不再提供绿色便携版，仅保留 NSIS 安装包。

@@ -1,6 +1,6 @@
 # Metona Ollama Desktop — 开发规范

-> 版本: v1-stable-release | 更新: 2026-04-24 | 维护: 项目团队
+> 版本: v1.1.0 | 更新: 2026-04-24 | 维护: 项目团队

 ---

@@ -1,12 +1,12 @@
 {
  "name": "metona-ollama-desktop",
-  "version": "1.0.0",
+  "version": "1.1.0",
  "lockfileVersion": 3,
  "requires": true,
  "packages": {
    "": {
      "name": "metona-ollama-desktop",
-      "version": "1.0.0",
+      "version": "1.1.0",
      "license": "MIT",
      "dependencies": {
        "sql.js": "^1.11.0"
@@ -1867,7 +1867,7 @@
      "license": "Python-2.0"
    },
    "node_modules/assert-plus": {
-      "version": "v1-stable-release",
+      "version": "1.0.0",
      "resolved": "https://registry.npmmirror.com/assert-plus/-/assert-plus-1.0.0.tgz",
      "integrity": "sha512-NfJ4UzBCcQGLDlQq7nHxH+tv3kyZ0hHQqF5BO6J7tNJeP5do1llPr8dZ8zHonfhAu0PHAdMkSo+8o0wxg9lZWw==",
      "dev": true,
@@ -1913,7 +1913,7 @@
      "license": "MIT"
    },
    "node_modules/at-least-node": {
-      "version": "v1-stable-release",
+      "version": "1.0.0",
      "resolved": "https://registry.npmmirror.com/at-least-node/-/at-least-node-1.0.0.tgz",
      "integrity": "sha512-+q/t7Ekv1EDY2l6Gda6LLiX14rU9TV20Wa3ofeQmwPFZbOMo9DXrLbOjFaaclkXKWidIaopwAObQDqwWtGUjqg==",
      "dev": true,
@@ -1,6 +1,6 @@
 {
  "name": "metona-ollama-desktop",
-  "version": "1.0.0",
+  "version": "1.1.0",
  "description": "Metona Ollama - TypeScript + Electron 桌面 AI 聊天客户端",
  "main": "dist/main/main.js",
  "author": "thzxx",
@@ -48,7 +48,7 @@
      "requestedExecutionLevel": "asInvoker"
    },
    "nsis": {
-      "artifactName": "Metona Ollama Setup v1-stable-release.${ext}",
+      "artifactName": "Metona Ollama Setup v1.1.0.${ext}",
      "oneClick": false,
      "allowToChangeInstallationDirectory": true,
      "createDesktopShortcut": true,
@@ -1,4 +1,4 @@
-# Metona Ollama Desktop v1-stable-release
+# Metona Ollama Desktop v1.1.0

 > 本地 Ollama AI 桌面客户端 — 首个稳定版本

@@ -135,7 +135,7 @@ Metona Ollama Desktop 是一个基于 **Ollama** 的本地 AI 桌面客户端，

 ## 安装说明

-1. 下载 `Metona Ollama Setup v1-stable-release.exe`
+1. 下载 `Metona Ollama Setup v1.1.0.exe`
 2. 运行安装程序，可自定义安装目录
 3. 确保本地已安装并运行 [Ollama](https://ollama.com)
 4. 启动 Metona Ollama，在设置中配置 Ollama 服务地址
@@ -101,7 +101,7 @@ export function createMenu(): void {
            dialog.showMessageBox(mainWindow!, {
              type: 'info',
              title: '关于 Metona Ollama',
-              message: 'Metona Ollama Desktop v3.2.2',
+              message: 'Metona Ollama Desktop v1.1.0',
              detail: 'TypeScript + Electron Ollama AI 聊天客户端\n\nhttps://gitee.com/thzxx/metona-ollama',
              icon: getIconPath()
            });
@@ -6,6 +6,7 @@ import { state, KEYS } from '../state/state.js';
 import { logError, logSession } from '../services/log-service.js';
 import { marked } from '../utils/marked-config.js';
 import { escapeHtml, formatTime } from '../utils/utils.js';
+import { estimateTokens } from '../services/context-manager.js';
 import type { ChatSession, ChatMessage, ToolCallRecord } from '../types.js';

 function getFileIcon(filename: string): string {
@@ -30,6 +31,31 @@ let scrollBtnEl: HTMLElement;
 let autoScroll = true;
 let currentPlaceholder: HTMLElement | null = null;

+/** 渲染可折叠系统提示词卡片 */
+let _sysPromptIdCounter = 0;
+function renderSystemPromptCard(sysPrompt: string): string {
+  const id = `sysprompt_${++_sysPromptIdCounter}`;
+  const tokenEstimate = estimateTokens(sysPrompt);
+  const preview = sysPrompt.length > 120 ? sysPrompt.slice(0, 120).replace(/\n/g, ' ') + '…' : sysPrompt.replace(/\n/g, ' ');
+  return `<div class="system-prompt-card">
+    <div class="system-prompt-header" onclick="
+      var el=document.getElementById('${id}');
+      var chev=this.querySelector('.sp-chevron');
+      el.style.display=el.style.display==='none'?'':'none';
+      chev.classList.toggle('sp-expanded');
+    ">
+      <span class="sp-icon">📋</span>
+      <span class="sp-title">系统提示词</span>
+      <span class="sp-tokens">~${tokenEstimate} tokens</span>
+      <span class="sp-preview">${escapeHtml(preview)}</span>
+      <svg class="sp-chevron" width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><polyline points="6 9 12 15 18 9"/></svg>
+    </div>
+    <div class="system-prompt-body" id="${id}" style="display:none;">
+      <pre>${escapeHtml(sysPrompt)}</pre>
+    </div>
+  </div>`;
+}
+
 export function initChatArea(): void {
  chatAreaEl = document.querySelector('#chatArea')!;
  messagesContainerEl = document.querySelector('#messagesContainer')!;
@@ -118,6 +144,12 @@ export function appendMessageDOM(msg: ChatMessage, index: number): void {
  const safeContent = (msg.content != null) ? String(msg.content) : '';

  if (msg.role === 'assistant') {
+    // ── 系统提示词折叠卡片（每条助理消息顶部展示）──
+    const sysPrompt = state.get<string>('_lastSystemPrompt', '');
+    if (sysPrompt) {
+      contentHtml += renderSystemPromptCard(sysPrompt);
+    }
+
    if (msg.think) {
      contentHtml += `
        <div class="think-block">
@@ -273,6 +305,21 @@ export function updateLastAssistantMessage(
    const loadingText = lastMsg.querySelector('.loading-text');
    if (loadingDots) loadingDots.remove();
    if (loadingText) loadingText.remove();
+
+    // 流式消息首次转正：注入系统提示词折叠卡片
+    const sysPrompt = state.get<string>('_lastSystemPrompt', '');
+    if (sysPrompt && !lastMsg.querySelector('.system-prompt-card')) {
+      const msgBody = lastMsg.querySelector('.msg-body');
+      const tempDiv = document.createElement('div');
+      tempDiv.innerHTML = renderSystemPromptCard(sysPrompt);
+      const card = tempDiv.firstElementChild!;
+      const thinkBlock = msgBody?.querySelector('.think-block');
+      if (thinkBlock) {
+        msgBody!.insertBefore(card, thinkBlock);
+      } else {
+        msgBody!.insertBefore(card, msgBody!.firstChild);
+      }
+    }
  }

  if (model) {
@@ -12,7 +12,7 @@ import {
  appendToolCallCardToPlaceholder, updateToolCallCardInPlaceholder, resetCurrentPlaceholder
 } from './chat-area.js';
 import { showToast } from './toast.js';
-import { addToolCard, updateToolCard, clearToolCardsExternal, clearTerminalExternal } from './workspace-panel.js';
+import { addToolCard, updateToolCard, clearToolCardsExternal, clearTerminalExternal, getWorkspaceDirPath } from './workspace-panel.js';
 import { ChatDB } from '../db/chat-db.js';
 import { OllamaAPI } from '../api/ollama.js';
 import { runAgentLoop } from '../services/agent-engine.js';
@@ -94,9 +94,7 @@ export function initInputArea(): void {
    const bridge = getBridge();
    if (!bridge) return;
    const paths = await bridge.dialog.openFile({
-      filters: [{ name: '图片', extensions: ['png', 'jpg', 'jpeg', 'gif', 'webp', 'bmp', 'svg'] }],
-      properties: ['openFile'],
-      multiple: false
+      filters: [{ name: '图片', extensions: ['png', 'jpg', 'jpeg', 'gif', 'webp', 'bmp', 'svg'] }]
    });
    if (!paths || paths.length === 0) return;
    await handleImagePaths(paths);
@@ -141,7 +139,7 @@ async function handleImagePaths(paths: string[]): Promise<void> {
  for (const filePath of paths) {
    const name = filePath.split(/[/\\]/).pop() || filePath;
    try {
-      const result = await bridge.fs.readFileBase64(filePath);
+      const result = await bridge!.fs.readFileBase64(filePath);
      if (!result.success) {
        appendSystemMessage(`❌ 读取 ${name} 失败: ${result.error}`);
        continue;
@@ -186,7 +184,7 @@ async function handleTextFilePaths(paths: string[]): Promise<void> {
      continue;
    }
    try {
-      const result = await bridge.fs.readFile(filePath);
+      const result = await bridge!.fs.readFile(filePath);
      if (!result.success) {
        appendSystemMessage(`❌ 读取 ${name} 失败: ${result.error}`);
        continue;
@@ -308,7 +306,7 @@ async function handleRetry(): Promise<void> {
              ...(retryThinkContent && { think: retryThinkContent }),
              ...(toolCallRecords?.length && { toolCalls: toolCallRecords }),
            };
-            state.update(KEYS.CURRENT_SESSION, (s: ChatSession | null) => ({
+            state.update(KEYS.CURRENT_SESSION, (s: any) => ({
              ...s, messages: [...s.messages, prevMsg], updatedAt: Date.now()
            }));
            renderMessages();
@@ -353,7 +351,7 @@ async function handleRetry(): Promise<void> {
                ...(loopStats?.prompt_eval_count && { prompt_eval_count: loopStats.prompt_eval_count }),
                ...(loopStats?.total_duration && { total_duration: loopStats.total_duration }),
              };
-              state.update(KEYS.CURRENT_SESSION, (s: ChatSession | null) => ({
+              state.update(KEYS.CURRENT_SESSION, (s: any) => ({
                ...s, messages: [...s.messages, lastMsg], updatedAt: Date.now()
              }));
            }
@@ -367,7 +365,7 @@ async function handleRetry(): Promise<void> {
              ...(loopStats?.prompt_eval_count && { prompt_eval_count: loopStats.prompt_eval_count }),
              ...(loopStats?.total_duration && { total_duration: loopStats.total_duration }),
            };
-            state.update(KEYS.CURRENT_SESSION, (s: ChatSession | null) => ({
+            state.update(KEYS.CURRENT_SESSION, (s: any) => ({
              ...s, messages: [...s.messages, assistantMsg], updatedAt: Date.now()
            }));
            updateLastAssistantMessage(finalContent, retryThinkContent || null, loopStats || null, undefined, Date.now(), toolRecords?.length ? toolRecords : undefined);
@@ -644,7 +642,7 @@ export async function sendMessage(): Promise<void> {
  }

  const isFirstMsg = currentSession.messages.length === 0;
-  state.update(KEYS.CURRENT_SESSION, (session: ChatSession | null) => ({
+  state.update(KEYS.CURRENT_SESSION, (session: any) => ({
    ...session,
    ...(isFirstMsg && {
      title: truncate(text || (pendingFiles.length > 0 ? `[文件: ${pendingFiles.map(f => f.name).join(', ')}]` : '[图片消息]'), 30),
@@ -682,7 +680,7 @@ export async function sendMessage(): Promise<void> {

  let assistantContent = '';
  let thinkContent = '';
-  let finalStats: { eval_count?: number; prompt_eval_count?: number; total_duration?: number } | null = null;
+  let finalStats: any = null;
  let modelName = '';

  try {
@@ -697,6 +695,20 @@ export async function sendMessage(): Promise<void> {
    };
    if (numCtx) chatParams.options = { num_ctx: numCtx, temperature };

+    // ── 扫描工作空间 SOUL.md ──
+    let soulContent = '';
+    const workspaceDir = getWorkspaceDirPath();
+    if (workspaceDir) {
+      try {
+        const soulResult = await window.metonaDesktop?.workspace.readFile(
+          workspaceDir.replace(/\/+$/, '') + '/SOUL.md'
+        );
+        if (soulResult?.success && soulResult.content) {
+          soulContent = `[SOUL.md] ${soulResult.content}`;
+        }
+      } catch { /* SOUL.md 不存在，静默跳过 */ }
+    }
+
    // ── Agent 记忆注入 ──
    if (isMemoryEnabled()) {
      const userMessage = text || (msgsToAdd.find(m => m.role === 'user')?.content || '');
@@ -704,7 +716,7 @@ export async function sendMessage(): Promise<void> {
        const relevantMemories = searchMemories(userMessage, 6);
        if (relevantMemories.length > 0) {
          const memoryContext = buildMemoryContext(relevantMemories);
-          chatParams.system = memoryContext;
+          chatParams.system = (soulContent ? soulContent + '\n\n' : '') + memoryContext;
          for (const m of relevantMemories) {
            await markMemoryUsed(m.id);
          }
@@ -712,6 +724,12 @@ export async function sendMessage(): Promise<void> {
        }
      }
    }
+    if (soulContent && !chatParams.system) {
+      chatParams.system = soulContent;
+    }
+
+    // 存入 state 供 chat-area 渲染系统提示词卡片
+    state.set('_lastSystemPrompt', chatParams.system || '');

    await api.chatStream(chatParams as any, (chunk: OllamaStreamChunk) => {
      if (chunk.message) {
@@ -752,9 +770,9 @@ export async function sendMessage(): Promise<void> {
      timestamp: Date.now(),
      ...(modelName && { model: modelName }),
      ...(thinkContent && { think: thinkContent }),
-      ...(finalStats && { eval_count: finalStats.eval_count, prompt_eval_count: finalStats.prompt_eval_count, total_duration: finalStats.total_duration })
+      ...((finalStats || {}) as any)
    };
-    state.update(KEYS.CURRENT_SESSION, (session: ChatSession | null) => ({
+    state.update(KEYS.CURRENT_SESSION, (session: any) => ({
      ...session,
      messages: [...session.messages, assistantMsg],
      updatedAt: Date.now()
@@ -800,7 +818,7 @@ export async function sendMessage(): Promise<void> {
        ...(thinkContent && { think: thinkContent }),
        stopped: true
      };
-      state.update(KEYS.CURRENT_SESSION, (session: ChatSession | null) => ({
+      state.update(KEYS.CURRENT_SESSION, (session: any) => ({
        ...session,
        messages: [...session.messages, partialMsg],
        updatedAt: Date.now()
@@ -829,7 +847,7 @@ export async function sendMessage(): Promise<void> {
    const contentDiv = placeholder?.querySelector('.msg-content');

    if (assistantContent && contentDiv) {
-      state.update(KEYS.CURRENT_SESSION, (session: ChatSession | null) => ({
+      state.update(KEYS.CURRENT_SESSION, (session: any) => ({
        ...session,
        messages: [...session.messages, { role: 'assistant', content: assistantContent + '\n\n`[已中断]`', timestamp: Date.now() } as ChatMessage],
        updatedAt: Date.now()
@@ -864,14 +882,14 @@ export async function sendMessage(): Promise<void> {
 /** 更新当前会话消息中最近一个 'running' 状态的同名工具记录 */
 function updateMessageToolRecord(toolName: string, status: 'success' | 'error', result: Record<string, unknown>): void {
  let updated = false;
-  state.update(KEYS.CURRENT_SESSION, (session: ChatSession | null) => {
+  state.update(KEYS.CURRENT_SESSION, (session: any) => {
    if (!session) return session;
    const messages = [...session.messages];
    // 从后往前找最近一条包含该工具 running 状态记录的消息
    for (let i = messages.length - 1; i >= 0; i--) {
      const msg = messages[i];
      if (!msg.toolCalls) continue;
-      const idx = msg.toolCalls.findIndex(tc => tc.name === toolName && tc.status === 'running');
+      const idx = msg.toolCalls.findIndex((tc: any) => tc.name === toolName && tc.status === 'running');
      if (idx !== -1) {
        const updatedRecords = [...msg.toolCalls];
        updatedRecords[idx] = { ...updatedRecords[idx], status, result };
@@ -919,7 +937,7 @@ async function sendMessageWithAgentLoop(text: string, currentSession: ChatSessio
  }

  const isFirstMsg = currentSession.messages.length === 0;
-  state.update(KEYS.CURRENT_SESSION, (session: ChatSession | null) => ({
+  state.update(KEYS.CURRENT_SESSION, (session: any) => ({
    ...session,
    ...(isFirstMsg && {
      title: truncate(text || (pendingFiles.length > 0 ? `[文件: ${pendingFiles.map(f => f.name).join(', ')}]` : '[图片消息]'), 30),
@@ -1009,7 +1027,7 @@ async function sendMessageWithAgentLoop(text: string, currentSession: ChatSessio
            ...(iterationEvalCount > 0 && { eval_count: iterationEvalCount }),
            ...(iterationDuration > 0 && { total_duration: iterationDuration }),
          };
-          state.update(KEYS.CURRENT_SESSION, (session: ChatSession | null) => ({
+          state.update(KEYS.CURRENT_SESSION, (session: any) => ({
            ...session,
            messages: [...session.messages, prevMsg],
            updatedAt: Date.now()
@@ -1075,7 +1093,7 @@ async function sendMessageWithAgentLoop(text: string, currentSession: ChatSessio
              ...(loopStats?.prompt_eval_count && { prompt_eval_count: loopStats.prompt_eval_count }),
              ...(loopStats?.total_duration && { total_duration: loopStats.total_duration }),
            };
-            state.update(KEYS.CURRENT_SESSION, (session: ChatSession | null) => ({
+            state.update(KEYS.CURRENT_SESSION, (session: any) => ({
              ...session,
              messages: [...session.messages, lastMsg],
              updatedAt: Date.now()
@@ -1094,7 +1112,7 @@ async function sendMessageWithAgentLoop(text: string, currentSession: ChatSessio
            ...(loopStats?.prompt_eval_count && { prompt_eval_count: loopStats.prompt_eval_count }),
            ...(loopStats?.total_duration && { total_duration: loopStats.total_duration }),
          };
-          state.update(KEYS.CURRENT_SESSION, (session: ChatSession | null) => ({
+          state.update(KEYS.CURRENT_SESSION, (session: any) => ({
            ...session,
            messages: [...session.messages, assistantMsg],
            updatedAt: Date.now()
@@ -1133,7 +1151,7 @@ async function sendMessageWithAgentLoop(text: string, currentSession: ChatSessio
        ...(abortToolRecords?.length && { toolCalls: abortToolRecords }),
        stopped: true
      };
-      state.update(KEYS.CURRENT_SESSION, (session: ChatSession | null) => ({
+      state.update(KEYS.CURRENT_SESSION, (session: any) => ({
        ...session,
        messages: [...session.messages, partialMsg],
        updatedAt: Date.now()
@@ -122,11 +122,22 @@ async function filterEmbedModels(models: Array<{ name: string }>): Promise<void>
      const info = await api.showModel(m.name);
      const caps = info.capabilities || [];
      const isCompletion = caps.includes('completion');
+
+      // 从 model_info 提取上下文长度
+      let contextLength = 2048;
+      for (const key of Object.keys(info.model_info || {})) {
+        if (key.endsWith('.context_length')) {
+          contextLength = Number(info.model_info![key]) || 2048;
+          break;
+        }
+      }
+
      modelCapabilityCache.set(m.name, {
        think: caps.includes('thinking'),
        vision: caps.includes('vision'),
        tools: caps.includes('tools'),
        completion: isCompletion,
+        contextLength,
      });
      if (!isCompletion) {
        const opt = modelSelectEl.querySelector(`option[value="${CSS.escape(m.name)}"]`);
@@ -149,7 +160,11 @@ async function filterEmbedModels(models: Array<{ name: string }>): Promise<void>

 async function checkModelCapability(modelName: string): Promise<void> {
  if (modelCapabilityCache.has(modelName)) {
-    applyCapability(modelName, modelCapabilityCache.get(modelName)!);
+    const cached = modelCapabilityCache.get(modelName)!;
+    state.set(KEYS.NUM_CTX, cached.contextLength);
+    const displayEl = document.querySelector('#displayNumCtx');
+    if (displayEl) displayEl.textContent = `${cached.contextLength.toLocaleString()} tokens`;
+    applyCapability(modelName, cached);
    return;
  }

@@ -159,18 +174,40 @@ async function checkModelCapability(modelName: string): Promise<void> {
  try {
    const info = await api.showModel(modelName);
    const capabilities = info.capabilities || [];
-    const caps = {
+
+    // 从 model_info 中提取模型实际上下文长度
+    let contextLength = 2048; // 默认保守值
+    const modelInfo = info.model_info || {};
+    for (const key of Object.keys(modelInfo)) {
+      if (key.endsWith('.context_length')) {
+        contextLength = Number(modelInfo[key]) || 2048;
+        break;
+      }
+    }
+
+    const caps: ModelCaps = {
      think: capabilities.includes('thinking'),
      vision: capabilities.includes('vision'),
      tools: capabilities.includes('tools'),
      completion: capabilities.includes('completion'),
+      contextLength,
    };
    modelCapabilityCache.set(modelName, caps);
-    logDebug(`能力检测: ${modelName}`, capabilities.join(', '));
+    logDebug(`能力检测: ${modelName}`, `ctx: ${contextLength}, ` + capabilities.join(', '));
+
+    // 自动设置上下文长度为模型实际支持的值
+    state.set(KEYS.NUM_CTX, contextLength);
+    const db = state.get<any>(KEYS.DB);
+    if (db) await db.saveSetting('numCtx', contextLength);
+
+    // 更新设置面板显示
+    const displayEl = document.querySelector('#displayNumCtx');
+    if (displayEl) displayEl.textContent = `${contextLength.toLocaleString()} tokens`;
+
    applyCapability(modelName, caps);
  } catch (err) {
    logWarn(`能力检测失败: ${modelName}`, (err as Error).message);
-    const fallback = { think: false, vision: false, tools: false, completion: true };
+    const fallback: ModelCaps = { think: false, vision: false, tools: false, completion: true, contextLength: 2048 };
    modelCapabilityCache.set(modelName, fallback);
    applyCapability(modelName, fallback);
  }
@@ -39,16 +39,6 @@ export function initSettingsModal(): void {
  }, 500);
  document.querySelector('#inputServerUrl')!.addEventListener('input', saveServerUrl);

-  const saveNumCtx = debounce(async () => {
-    const db = state.get<ChatDB | null>(KEYS.DB);
-    const val = (document.querySelector('#inputNumCtx') as HTMLInputElement).value.trim();
-    const numCtx = val ? parseInt(val) : 24576;
-    state.set(KEYS.NUM_CTX, numCtx);
-    if (db) await db.saveSetting('numCtx', numCtx);
-    logSetting('上下文长度', `${numCtx} tokens`);
-  }, 500);
-  document.querySelector('#inputNumCtx')!.addEventListener('input', saveNumCtx);
-
  const tempSlider = document.querySelector('#inputTemperature') as HTMLInputElement;
  const tempDisplay = document.querySelector('#tempValue')!;
  tempSlider.addEventListener('input', () => {
@@ -250,9 +240,9 @@ export function initSettingsModal(): void {
      const result = await bridge.workspace.setDir(newDir);
      if (result.success) {
        showToast('工作空间目录已更新', 'success');
-        logSetting('工作空间目录', result.dir);
+        logSetting('工作空间目录', result.dir!);
      } else {
-        showToast(`设置失败: ${result.error}`, 'error');
+        showToast(`设置失败: ${result.error!}`, 'error');
        // 恢复原值
        const info = await bridge.workspace.getDir();
        inputWorkspaceDir.value = info.dir;
@@ -401,7 +391,7 @@ async function importSessions(filePath: string): Promise<void> {

  try {
    const bridge = window.metonaDesktop;
-    const result = await bridge.fs.readFileBase64(filePath);
+    const result = await bridge!.fs.readFileBase64(filePath);
    if (!result.success) {
      showToast(`读取文件失败: ${result.error}`, 'error');
      return;
@@ -301,7 +301,7 @@ function renderBarChart(

 /** 渲染全局柱状图（按会话） */
 function renderGlobalChart(stats: AllSessionsTokenStats): void {
-  const chartEl = dashboardModalEl.querySelector('#tdChart')!;
+  const chartEl = dashboardModalEl.querySelector('#tdChart') as HTMLElement;
  const barsEl = chartEl.querySelector('.td-chart-bars') as HTMLElement | null;
  const savedScrollLeft = barsEl?.scrollLeft ?? 0;

@@ -329,7 +329,7 @@ function renderGlobalChart(stats: AllSessionsTokenStats): void {

 /** 渲染全局明细表格 */
 function renderGlobalTable(stats: AllSessionsTokenStats): void {
-  const tableTitleEl = dashboardModalEl.querySelector('#tdTableTitle');
+  const tableTitleEl = dashboardModalEl.querySelector('#tdTableTitle') as HTMLElement | null;
  if (tableTitleEl) tableTitleEl.textContent = '📋 会话汇总明细';

  // 动态更新表头
@@ -406,7 +406,7 @@ function renderSessionView(): void {
  `;

  // 柱状图
-  const chartEl = dashboardModalEl.querySelector('#tdChart')!;
+  const chartEl = dashboardModalEl.querySelector('#tdChart') as HTMLElement;
  const barsEl = chartEl.querySelector('.td-chart-bars') as HTMLElement | null;
  const savedScrollLeft = barsEl?.scrollLeft ?? 0;

@@ -423,7 +423,7 @@ function renderSessionView(): void {
  renderBarChart(chartEl, chartData, '按轮次', undefined, savedScrollLeft);

  // 表格标题
-  const tableTitleEl = dashboardModalEl.querySelector('#tdTableTitle');
+  const tableTitleEl = dashboardModalEl.querySelector('#tdTableTitle') as HTMLElement | null;
  if (tableTitleEl) tableTitleEl.textContent = '📋 轮次明细（按时间倒序）';

  // 动态更新表头
@@ -7,7 +7,7 @@ import { state, KEYS } from '../state/state.js';
 import { setToolEnabled, setRunCommandMode } from '../services/tool-registry.js';
 import { showToast } from './toast.js';
 import { logInfo, logWarn } from '../services/log-service.js';
-import type { ChatDB } from '../types.js';
+import type { ChatDB } from '../db/chat-db.js';

 let modalEl: HTMLElement;

@@ -10,9 +10,9 @@ function isDesktop(): boolean {
  return !!window.metonaDesktop?.isDesktop;
 }

-/** 桌面端 DB 桥接 */
+/** 桌面端 DB 桥接（isDesktop 已检查，安全断言非空） */
 function dbBridge() {
-  return window.metonaDesktop?.db;
+  return window.metonaDesktop!.db!;
 }

 export class ChatDB {
@@ -201,7 +201,7 @@ export class ChatDB {

  async getSetting<T = unknown>(key: string, defaultValue: T | null = null): Promise<T> {
    if (isDesktop()) {
-      return dbBridge().getSetting(key, defaultValue);
+      return dbBridge().getSetting(key, defaultValue) as Promise<T>;
    }
    return this._idbGetSetting(key, defaultValue);
  }
@@ -28,7 +28,7 @@
        <div class="header-left">
            <span class="logo">🦙</span>
            <span class="app-title">Metona Ollama</span>
-            <span class="app-version">v1-stable-release</span>
+            <span class="app-version">v1.1.0</span>
            <button class="icon-btn help-btn" id="btnHelp" title="使用帮助">
                <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2">
                    <circle cx="12" cy="12" r="10"/><path d="M9.09 9a3 3 0 0 1 5.83 1c0 2-3 3-3 3"/>
@@ -272,12 +272,11 @@
                </div>
            </div>
            <div class="setting-group">
-                <label class="setting-label">上下文长度（num_ctx）</label>
+                <label class="setting-label">上下文长度（自动检测）</label>
                <div style="display:flex;gap:8px;align-items:center;">
-                    <input class="setting-input" id="inputNumCtx" type="number" min="512" max="131072" step="512" value="24576" style="margin-bottom:0;">
-                    <span class="text-muted" style="white-space:nowrap;">tokens</span>
+                    <span class="setting-input" id="displayNumCtx" style="margin-bottom:0;display:flex;align-items:center;background:var(--bg-layer);padding:9px 12px;">24576 tokens</span>
                </div>
-                <p class="text-muted" style="margin-top:4px;font-size:11px;">常见值：2048 / 4096 / 8192 / 32768。值越大上下文越长，显存占用越高。</p>
+                <p class="text-muted" style="margin-top:4px;font-size:11px;">选择模型时自动从模型元数据中获取实际上下文长度。</p>
                <label class="setting-label" style="margin-top:12px;">温度（Temperature）：<span id="tempValue">0.7</span></label>
                <input type="range" id="inputTemperature" class="temp-slider" min="0" max="2" step="0.1" value="0.7">
                <p class="text-muted" style="margin-top:4px;font-size:11px;">低值更精确稳定，高值更有创意多样（0 = 确定性，2 = 最大随机）</p>
@@ -100,8 +100,8 @@ async function migrateIndexedDBToSQLite(db: ChatDB): Promise<void> {
        created_at: s.createdAt,
        updated_at: s.updatedAt
      })),
-      messages: sessions.flatMap(s =>
-        s.messages.map((m: Record<string, unknown>, idx: number) => ({
+      messages: (sessions as any[]).flatMap((s: any) =>
+        (s.messages || []).map((m: any, idx: number) => ({
          id: `${s.id}_msg_${idx}_${m.timestamp}`,
          session_id: s.id,
          role: m.role,
@@ -120,7 +120,7 @@ async function migrateIndexedDBToSQLite(db: ChatDB): Promise<void> {
        type: m.type,
        content: m.content,
        importance: m.importance || 5,
-        tags: m.tags?.length ? JSON.stringify(m.tags) : null,
+        tags: (m.tags as any[])?.length ? JSON.stringify(m.tags) : null,
        source: m.source || null,
        session_id: m.sessionId || null,
        use_count: m.useCount || 0,
@@ -354,9 +354,8 @@ async function init(): Promise<void> {
    state.set(KEYS.NUM_CTX, numCtx);
    state.set('temperature', temperature);

-    (document.querySelector('#inputNumCtx') as HTMLInputElement).value = String(numCtx);
    (document.querySelector('#inputTemperature') as HTMLInputElement).value = String(temperature);
-    document.querySelector('#tempValue')!.textContent = parseFloat(temperature).toFixed(1);
+    document.querySelector('#tempValue')!.textContent = parseFloat(String(temperature)).toFixed(1);

    // ── v5.1.1 Agent 最大轮数 ──
    const maxTurns = await db.getSetting('maxTurns', 85);
@@ -18,14 +18,15 @@ import { showToast } from '../components/toast.js';
 import { logInfo, logWarn, logSuccess, logError, logToolStart, logToolResult, logAgentLoop, logModelResponse } from './log-service.js';
 import { getWorkspaceDirPath } from '../components/workspace-panel.js';
 import { generateId } from '../utils/utils.js';
-import { buildContext, estimateTokens, shouldAutoCompress, compressWithLLM, AUTO_COMPRESS_THRESHOLD } from './context-manager.js';
+import { buildContext, estimateTokens, shouldAutoCompress, compressWithLLM, AUTO_COMPRESS_THRESHOLD, recordActualTokens } from './context-manager.js';
 import type {
  OllamaMessage,
  OllamaStreamChunk,
  ToolCall,
  ToolResult,
  ToolCallRecord,
-  TraceEntry
+  TraceEntry,
+  ChatSession
 } from '../types.js';

 const MAX_RETRIES = 2;              // 工具错误自动重试次数
@@ -173,7 +174,7 @@ async function inferUserProfile(userMsg: string, assistantMsg: string): Promise<
  const bridge = window.metonaDesktop;
  if (!bridge?.db?.getSetting) return;

-  const profile: Record<string, unknown> = await bridge.db.getSetting('user_profile', {}) || {};
+  const profile: Record<string, any> = await bridge.db.getSetting('user_profile', {}) || {};

  // 技术栈检测
  const techPatterns: Record<string, RegExp> = {
@@ -363,7 +364,7 @@ export interface AgentCallbacks {
 }

 /** 保存执行轨迹到 SQLite */
-async function saveTrace(trace: Omit<TraceEntry, 'id' | 'createdAt'>): Promise<void> {
+async function saveTrace(trace: Record<string, any>): Promise<void> {
  try {
    const bridge = window.metonaDesktop;
    if (!bridge?.db) return;
@@ -390,7 +391,7 @@ export async function runAgentLoop(
 ): Promise<void> {
  const api = state.get<OllamaAPI>(KEYS.API);
  const model = state.get<string>('_defaultModel', '');
-  const currentSession = state.get(KEYS.CURRENT_SESSION);
+  const currentSession = state.get<ChatSession | null>(KEYS.CURRENT_SESSION);
  const sessionId = currentSession?.id || 'unknown';

  if (!api || !model) {
@@ -411,6 +412,24 @@ export async function runAgentLoop(
  const messages: OllamaMessage[] = [];
  let systemPromptParts: string[] = [];

+  // ── 扫描工作空间 SOUL.md ──
+  const workspaceDir = getWorkspaceDirPath();
+  if (workspaceDir) {
+    try {
+      const soulResult = await window.metonaDesktop?.workspace.readFile(
+        workspaceDir.replace(/\/+$/, '') + '/SOUL.md'
+      );
+      if (soulResult?.success && soulResult.content) {
+        // SOUL.md 注入为独立 system 消息，标记为不可压缩
+        messages.push({
+          role: 'system',
+          content: `[SOUL.md] ${soulResult.content}`,
+        });
+        logInfo('SOUL.md 已注入系统提示词', `${soulResult.lines || 0} 行`);
+      }
+    } catch { /* SOUL.md 不存在或读取失败，静默跳过 */ }
+  }
+
  // 注入记忆上下文
  if (isMemoryEnabled() && userContent) {
    const relevantMemories = searchMemories(userContent, 6);
@@ -423,7 +442,6 @@ export async function runAgentLoop(
  }

  // 注入工作空间上下文
-  const workspaceDir = getWorkspaceDirPath();
  if (workspaceDir) {
    systemPromptParts.push(`【工作空间】
 当前工作空间目录: ${workspaceDir}
@@ -447,7 +465,7 @@ export async function runAgentLoop(
  const db = state.get<ChatDB | null>(KEYS.DB);
  if (db) {
    try {
-      const userProfile = await db.getSetting('user_profile', null);
+      const userProfile = await db.getSetting('user_profile', null) as Record<string, any> | null;
      if (userProfile && typeof userProfile === 'object') {
        const parts: string[] = [];
        if (userProfile.tech_stack?.length) parts.push(`技术栈: ${userProfile.tech_stack.join(', ')}`);
@@ -475,6 +493,9 @@ export async function runAgentLoop(

  messages.push({ role: 'system', content: fullSystemPrompt });

+  // 将完整系统提示词存入 state，供 input-area 在助理消息上展示
+  state.set('_lastSystemPrompt', fullSystemPrompt);
+
  // 添加历史消息
  for (const msg of historyMessages) {
    messages.push({
@@ -503,7 +524,7 @@ export async function runAgentLoop(
    logInfo(`自动上下文压缩触发: tokens≈${estimateTokens(messages.map(m => m.content || '').join(''))} > ${Math.floor(numCtx * AUTO_COMPRESS_THRESHOLD)} (50% of ${numCtx})`);
    const compressAC = state.get<AbortController | null>(KEYS.ABORT_CONTROLLER) || new AbortController();
    try {
-      const compressed = await compressWithLLM(messages, api, model, { abortController: compressAC });
+      const compressed = await compressWithLLM(messages, api as any, model, { abortController: compressAC });
      if (compressed.length < messages.length || estimateTokens(compressed.map(m => m.content || '').join('')) < estimateTokens(messages.map(m => m.content || '').join(''))) {
        messages.length = 0;
        messages.push(...compressed);
@@ -633,10 +654,25 @@ export async function runAgentLoop(
    totalPromptEvalCount += loopPromptEvalCount;
    totalInferenceNs += loopInferenceNs;
    state.set('_currentEvalCount', totalEvalCount);
+
+    // Token 校准：用 Ollama 返回的实际计数修正估算器（在重置前保存本轮值）
+    const thisLoopEval = loopEvalCount;
+    const thisLoopPrompt = loopPromptEvalCount;
    // 重置本轮计数器（下一轮重新从 chunk 收集）
    loopEvalCount = 0;
    loopPromptEvalCount = 0;
    loopInferenceNs = 0;
+
+    // Token 校准：用 Ollama 返回的实际计数自动修正估算器
+    try {
+      if (thisLoopEval > 0 || thisLoopPrompt > 0) {
+        const estimatedThisLoop = estimateTokens(messages.map(m => m.content || '').join(''));
+        if (estimatedThisLoop > 0) {
+          recordActualTokens(thisLoopPrompt, thisLoopEval, estimatedThisLoop);
+        }
+      }
+    } catch { /* 校准失败不阻塞主流程 */ }
+
    } catch (err) {
      if (abortController.signal.aborted) {
        logInfo('流式调用已中止');
@@ -722,7 +758,7 @@ export async function runAgentLoop(
      // v4.2 自动提取技能
      if (allToolRecords.length >= 2) {
        try {
-          await extractSkillsFromToolRecords(allToolRecords, userContent, currentSession?.title || '');
+          await extractSkillsFromToolRecords(allToolRecords, userContent, (currentSession as ChatSession)?.title || '');
        } catch { /* 不阻塞 */ }
      }

@@ -7,7 +7,31 @@
 import type { OllamaMessage, OllamaStreamChunk } from '../types.js';
 import { logInfo, logWarn, logSuccess, logError } from './log-service.js';

-/** 粗略估算 token 数 */
+// ── Token 校准系统 ──
+
+/** 校准比例：actualTokens / estimatedTokens，基于 Ollama 返回的实际计数动态修正 */
+let _tokenCalibrationRatio = 1.0;
+let _calibrationSamples = 0;
+const MIN_CALIBRATION_SAMPLES = 3;
+
+/**
+ * 记录 Ollama 返回的实际 token 计数，用于校准估算器。
+ * 在 agent-engine.ts 每轮流式完成后调用。
+ * @param actualInputTokens Ollama 返回的 prompt_eval_count
+ * @param actualOutputTokens Ollama 返回的 eval_count
+ * @param estimatedTokens 本轮消息调用 estimateTokens 的合计值
+ */
+export function recordActualTokens(actualInputTokens: number, actualOutputTokens: number, estimatedCount: number): void {
+  if (actualInputTokens <= 0 || estimatedCount <= 0) return;
+  const actualTotal = actualInputTokens + actualOutputTokens;
+  const sampleRatio = actualTotal / Math.max(1, estimatedCount);
+  // 指数移动平均，平滑异常值
+  const alpha = 0.3;
+  _tokenCalibrationRatio = _tokenCalibrationRatio * (1 - alpha) + sampleRatio * alpha;
+  _calibrationSamples++;
+}
+
+/** 估算 token 数（自动使用校准后的比例） */
 export function estimateTokens(text: string): number {
  if (!text) return 0;
  // 中文按 1.5 字/token，英文按 4 字符/token
@@ -17,7 +41,17 @@ export function estimateTokens(text: string): number {
    if (/[\u4e00-\u9fff]/.test(ch)) chineseChars++;
    else otherChars++;
  }
-  return Math.ceil(chineseChars / 1.5 + otherChars / 4);
+  const raw = Math.ceil(chineseChars / 1.5 + otherChars / 4);
+  // 应用校准比例（仅在有足够样本后）
+  if (_calibrationSamples >= MIN_CALIBRATION_SAMPLES) {
+    return Math.ceil(raw * _tokenCalibrationRatio);
+  }
+  return raw;
+}
+
+/** 获取当前校准比例（供调试用） */
+export function getTokenCalibration(): { ratio: number; samples: number } {
+  return { ratio: _tokenCalibrationRatio, samples: _calibrationSamples };
 }

 /** 自动压缩阈值：当消息 token 占 context window 比例超过此值时触发自动压缩 */
@@ -27,6 +61,57 @@ export const AUTO_COMPRESS_THRESHOLD = 0.5;
 const COMPRESS_KEEP_HEAD = 2;
 const COMPRESS_KEEP_TAIL = 2;

+// ── 消息重要性评分（纯规则，不调用 LLM）──
+
+/** 重要性评分：0-10，越高越应该保留。SOUL.md 和规则记忆始终保留。 */
+export function scoreMessageImportance(msg: OllamaMessage): number {
+  let score = 5; // 默认中等
+
+  // SOUL.md 消息永远不可压缩
+  if (msg.content?.startsWith('[SOUL.md]')) return 10;
+
+  // 角色权重
+  if (msg.role === 'user') score += 2;      // 用户消息最重要
+  if (msg.role === 'system') score -= 2;    // 系统消息通常可压缩
+
+  // 已压缩标记 → 低权重（避免压缩产物堆积）
+  if (msg.compressed) score = 1;
+
+  const content = msg.content || '';
+
+  // 关键词检测
+  const highValuePatterns = [/路径|目录|path|file|config|配置|命令|command|exec/i,
+    /错误|error|失败|fail|bug|fix|修复|解决/i,
+    /版本|version|API|http|url|端口|port|localhost/i,
+    /记住|保存|memory|偏好|偏好|规则|rule/i,
+    /完成|done|✓|success|成功|结果|result/i,
+    /项目|project|工作空间|workspace|git|repo|仓库/i,
+  ];
+  const lowValuePatterns = [/好的|明白|ok|知道了|嗯|哦|好/i,
+    /继续|请|帮我|可以吗/i,
+    /谢谢|感谢|不客气/i,
+  ];
+
+  for (const p of highValuePatterns) {
+    if (p.test(content)) { score += 1; break; }
+  }
+  for (const p of lowValuePatterns) {
+    if (p.test(content) && content.length < 100) { score -= 2; break; }
+  }
+
+  // 工具调用 → 高价值
+  if (msg.tool_calls?.length) score += 2;
+
+  // 长度加分：长消息通常包含更多信息
+  if (content.length > 500) score += 1;
+  if (content.length > 2000) score += 1;
+
+  // 图像附件 → 中等价值（大但语义密度低）
+  if (msg.images?.length) score -= 1;
+
+  return Math.max(1, Math.min(10, score));
+}
+
 export interface ContextBuildOptions {
  /** 滑动窗口大小（最近 N 条消息完整保留） */
  windowSize?: number;
@@ -127,9 +212,28 @@ export function shouldAutoCompress(messages: OllamaMessage[], numCtx: number): b
  return totalTokens > threshold;
 }

+// ── 结构化压缩 ──
+
+/** 结构化摘要 */
+export interface StructuredSummary {
+  /** 讨论的主题 */
+  topics: string[];
+  /** 已做出的决定 */
+  decisions: string[];
+  /** 未完成的待办事项 */
+  pendingTasks: string[];
+  /** 发现的约束/规则 */
+  constraints: string[];
+  /** 关键知识点（跨会话有价值的信息） */
+  knowledge: string[];
+  /** 工具调用结果摘要 */
+  toolResults: string[];
+}
+
 /**
- * LLM 摘要压缩：调用模型对中间消息生成摘要
- * 保留首尾各 keepHead/keepTail 条消息，中间用 LLM 摘要替换
+ * LLM 摘要压缩：调用模型对中间消息生成结构化 JSON 摘要。
+ * 保留首尾各 keepHead/keepTail 条消息，中间用 JSON 摘要替换。
+ * v6.0: 输出结构化 JSON，支持增量合并。
 *
 * @returns 压缩后的消息列表（包含 compressed 标记的摘要消息）
 */
@@ -182,14 +286,14 @@ export async function compressWithLLM(

  logInfo(`上下文压缩: 开始 LLM 摘要，${uncompressedMiddle.length} 条消息待压缩`);

-  let summary = '';
+  let summaryJson = '';
  try {
    await api.chatStream(
      {
        model,
        messages: [{
          role: 'user',
-          content: `请将以下对话摘要为一段简洁的上下文总结。保留关键信息：讨论的主题、得出的结论、用户的偏好、未完成的任务、关键工具调用结果。用中文输出，不超过 ${maxSummaryTokens} 字。\n\n对话记录：\n${conversationText}`
+          content: `请将以下对话摘要为结构化 JSON。保留关键信息，用中文输出。严格按此 JSON 格式返回（不要输出其他内容）：\n\n{\n  "topics": ["讨论的 2-4 个核心主题"],\n  "decisions": ["已做出的决定（最多 3 条）"],\n  "pendingTasks": ["尚未完成的任务（最多 3 条）"],\n  "constraints": ["发现的约束/规则/偏好（最多 3 条）"],\n  "knowledge": ["跨会话有价值的长期知识点（最多 3 条）"],\n  "toolResults": ["关键工具调用结果摘要（最多 3 条）"]\n}\n\n对话记录：\n${conversationText}`
        }],
        stream: true,
        think: false,
@@ -197,7 +301,7 @@ export async function compressWithLLM(
      },
      (chunk: OllamaStreamChunk) => {
        if (chunk.message?.content) {
-          summary += chunk.message.content;
+          summaryJson += chunk.message.content;
        }
      },
      options.abortController
@@ -211,15 +315,34 @@ export async function compressWithLLM(
    return messages;
  }

-  if (!summary.trim()) {
+  if (!summaryJson.trim()) {
    logWarn('上下文压缩: 模型未返回摘要内容');
    return messages;
  }

+  // 解析 JSON 摘要（容错：提取第一个 JSON 块或直接解析）
+  let parsed: StructuredSummary;
+  try {
+    const jsonMatch = summaryJson.match(/\{[\s\S]*"topics"[\s\S]*\}/);
+    parsed = jsonMatch ? JSON.parse(jsonMatch[0]) : JSON.parse(summaryJson);
+  } catch {
+    logWarn('上下文压缩: JSON 解析失败，使用纯文本摘要');
+    parsed = { topics: [summaryJson.slice(0, 100)], decisions: [], pendingTasks: [], constraints: [], knowledge: [], toolResults: [] };
+  }
+
+  // 构建结构化摘要消息文本
+  const parts: string[] = [];
+  if (parsed.topics.length) parts.push(`📌 主题: ${parsed.topics.join('；')}`);
+  if (parsed.decisions.length) parts.push(`✅ 决策: ${parsed.decisions.join('；')}`);
+  if (parsed.pendingTasks.length) parts.push(`⏳ 待办: ${parsed.pendingTasks.join('；')}`);
+  if (parsed.constraints.length) parts.push(`📏 约束: ${parsed.constraints.join('；')}`);
+  if (parsed.knowledge.length) parts.push(`🧠 知识点: ${parsed.knowledge.join('；')}`);
+  if (parsed.toolResults.length) parts.push(`🔧 工具结果: ${parsed.toolResults.join('；')}`);
+
  // 构建压缩后的摘要消息
  const summaryMsg: OllamaMessage = {
    role: 'system',
-    content: `📋 以下是对之前对话的摘要（已压缩 ${uncompressedMiddle.length} 条消息）：\n\n${summary}`,
+    content: `📋 以下是对之前对话的摘要（已压缩 ${uncompressedMiddle.length} 条消息）：\n\n${parts.join('\n')}`,
    compressed: true
  };

@@ -243,7 +366,8 @@ export async function compressWithLLM(
 }

 /**
- * 将较早的消息每 batchSize 条压缩为一段摘要（快速文本截取，不调用 LLM）
+ * 将较早的消息每 batchSize 条压缩为一段摘要。
+ * v6.0: 重要性高的消息保留更多内容，低价值消息激进截断。
 */
 function summarizeOlderMessages(messages: OllamaMessage[], batchSize: number): OllamaMessage[] {
  const summaries: OllamaMessage[] = [];
@@ -262,7 +386,7 @@ function summarizeOlderMessages(messages: OllamaMessage[], batchSize: number): O
 }

 /**
- * 快速摘要（不调用模型，直接提取关键信息）
+ * 快速摘要：重要性高的消息保留更多信息，低价值的激进截断
 */
 function createQuickSummary(messages: OllamaMessage[]): string {
  const parts: string[] = [];
@@ -270,7 +394,21 @@ function createQuickSummary(messages: OllamaMessage[]): string {
  for (const msg of messages) {
    const role = msg.role === 'user' ? '用户' : 'AI';
    const content = msg.content || '';
-    const preview = content.length > 100 ? content.slice(0, 100) + '...' : content;
+    const importance = scoreMessageImportance(msg);
+
+    let preview: string;
+    if (importance >= 8) {
+      // 高价值消息：保留 200 字
+      preview = content.length > 200 ? content.slice(0, 200) + '...' : content;
+    } else if (importance >= 5) {
+      // 中等价值：保留 100 字
+      preview = content.length > 100 ? content.slice(0, 100) + '...' : content;
+    } else {
+      // 低价值：仅保留 40 字或跳过
+      if (content.trim().length < 10) continue;
+      preview = content.length > 40 ? content.slice(0, 40) + '...' : content;
+    }
+
    if (preview.trim()) {
      parts.push(`${role}: ${preview}`);
    }
@@ -284,31 +422,46 @@ function createQuickSummary(messages: OllamaMessage[]): string {
 }

 /**
- * 根据 token 限制裁剪消息
+ * 根据 token 限制裁剪消息。
+ * v6.0: 按重要性评分决定保留顺序——重要性低的优先被丢弃。
 */
 function trimByTokenLimit(messages: OllamaMessage[], maxTokens: number): OllamaMessage[] {
-  let totalTokens = 0;
-  const result: OllamaMessage[] = [];
+  // 分离 system 和非 system 消息
+  const systemMsgs = messages.filter(m => m.role === 'system');
+  const nonSystemMsgs = messages.filter(m => m.role !== 'system');

-  for (let i = messages.length - 1; i >= 0; i--) {
-    const msg = messages[i];
-    const msgTokens = estimateTokens(msg.content || '') +
-      (msg.images ? msg.images.length * 100 : 0) +
-      (msg.tool_calls ? msg.tool_calls.length * 50 : 0);
+  if (nonSystemMsgs.length <= 4) return messages; // 消息太少不裁剪

-    // system 消息始终保留
-    if (msg.role === 'system') {
-      totalTokens += msgTokens;
-      result.unshift(msg);
-      continue;
-    }
+  // 计算每条消息的 token + 重要性
+  const scored = nonSystemMsgs.map(m => ({
+    msg: m,
+    tokens: estimateTokens(m.content || '') +
+      (m.images ? m.images.length * 100 : 0) +
+      (m.tool_calls ? m.tool_calls.length * 50 : 0),
+    importance: scoreMessageImportance(m),
+  }));

-    if (totalTokens + msgTokens > maxTokens && result.length > 2) {
-      break;
-    }
+  // system 消息的 token 消耗
+  const systemTokens = systemMsgs.reduce((sum, m) => sum + estimateTokens(m.content || ''), 0);
+  const availableTokens = maxTokens - systemTokens;

-    totalTokens += msgTokens;
-    result.unshift(msg);
+  // 按重要性降序排列，取能装下的最大数量
+  scored.sort((a, b) => b.importance - a.importance);
+
+  let usedTokens = 0;
+  const kept = new Set<number>(); // 保留的索引
+  for (let i = 0; i < scored.length; i++) {
+    if (usedTokens + scored[i].tokens > availableTokens && kept.size >= 2) break;
+    usedTokens += scored[i].tokens;
+    kept.add(i);
+  }
+
+  // 按原始顺序重建：system → 按重要性保留的非system
+  const result: OllamaMessage[] = [...systemMsgs];
+  // 使用原始顺序而非重要性顺序，保持对话的时间线
+  const keptSet = new Set(scored.filter((_, i) => kept.has(i)).map(s => s.msg));
+  for (const m of nonSystemMsgs) {
+    if (keptSet.has(m)) result.push(m);
  }

  return result;
@@ -9,9 +9,9 @@ const PBKDF2_ITERATIONS = 100000;

 async function deriveKeyBytes(salt: Uint8Array): Promise<Uint8Array> {
  const passBytes = new TextEncoder().encode(PASSPHRASE);
-  const keyMaterial = await crypto.subtle.importKey('raw', passBytes, 'PBKDF2', false, ['deriveKey']);
+  const keyMaterial = await (crypto.subtle as any).importKey('raw', passBytes, 'PBKDF2', false, ['deriveKey']);
  const key = await crypto.subtle.deriveKey(
-    { name: 'PBKDF2', salt, iterations: PBKDF2_ITERATIONS, hash: 'SHA-256' },
+    { name: 'PBKDF2', salt: salt as any, iterations: PBKDF2_ITERATIONS, hash: 'SHA-256' },
    keyMaterial,
    { name: 'AES-GCM', length: 256 },
    true,
@@ -27,7 +27,7 @@ export async function encryptData(data: unknown): Promise<Blob> {
  const iv = crypto.getRandomValues(new Uint8Array(12));
  const keyBytes = await deriveKeyBytes(salt);

-  const key = await crypto.subtle.importKey('raw', keyBytes, 'AES-GCM', false, ['encrypt']);
+  const key = await crypto.subtle.importKey('raw', keyBytes as any, 'AES-GCM', false, ['encrypt']);
  const ciphertext = await crypto.subtle.encrypt({ name: 'AES-GCM', iv }, key, json);
  const payload = new Uint8Array(ciphertext);

@@ -51,7 +51,7 @@ export async function decryptData(buffer: ArrayBuffer): Promise<unknown> {
    throw new Error('不支持的加密格式，请使用最新版本重新导出');
  }

-  const key = await crypto.subtle.importKey('raw', keyBytes, 'AES-GCM', false, ['decrypt']);
+  const key = await crypto.subtle.importKey('raw', keyBytes as any, 'AES-GCM', false, ['decrypt']);
  const decrypted = await crypto.subtle.decrypt({ name: 'AES-GCM', iv }, key, payload);
  const jsonBytes = new Uint8Array(decrypted);

@@ -17,7 +17,9 @@ import {
  getMemoryCollectionId
 } from './vector-memory.js';
 import { logMemory, logDebug, logWarn, logInfo } from './log-service.js';
-import type { MemoryEntry, MemorySearchResult, MemoryExtractionResult, ChatDB, OllamaAPI } from '../types.js';
+import type { MemoryEntry, MemorySearchResult, MemoryExtractionResult } from '../types.js';
+import type { ChatDB } from '../db/chat-db.js';
+import type { OllamaAPI } from '../api/ollama.js';

 const TYPE_ICONS: Record<string, string> = {
  fact: '📌',
@@ -686,10 +688,9 @@ ${conversationText.slice(0, 4000)}
    const response = await api.chat({
      model,
      messages: [{ role: 'user', content: extractPrompt }],
-      stream: false,
      think: false,
      options: { num_ctx: 8192, temperature: 0.1 }
-    });
+    } as any);

    const content = (response as { message?: { content?: string } })?.message?.content || '';
    const jsonMatch = content.match(new RegExp('```json\\s*([\\s\\S]*?)\\s*```')) || content.match(/\{[\s\S]*"entries"[\s\S]*\}/);
@@ -715,7 +716,7 @@ ${conversationText.slice(0, 4000)}
        importance: Math.min(10, Math.max(1, entry.importance || 5)),
        tags: entry.tags || [],
        source: '自动提取',
-        sessionId: currentSession?.id
+        sessionId: (currentSession as any)?.id
      });
      count++;
    }
@@ -64,7 +64,7 @@ export async function extractSkillsFromToolRecords(
  const chainNames = chain.map(s => s.name).join(' → ');

  // 检查是否已有完全相同的技能链
-  const existingSkills: Skill[] = await bridge.db.getAllSkills();
+  const existingSkills = await bridge.db.getAllSkills() as Skill[];
  const duplicate = existingSkills.find(s => {
    try {
      const existingChain: ToolChainStep[] = JSON.parse(s.tool_chain);
@@ -73,7 +73,29 @@ export async function extractSkillsFromToolRecords(
  });

  if (duplicate) {
-    // 已有相同技能，不重复创建
+    // v1.1: 重复链 → 更新参数提示（合并新知）+ 增加计数
+    try {
+      const existingChain: ToolChainStep[] = JSON.parse(duplicate.tool_chain);
+      let updated = false;
+      for (let i = 0; i < Math.min(chain.length, existingChain.length); i++) {
+        const newArgs = chain[i].args_hint.split(',').map(a => a.trim()).filter(Boolean);
+        const oldArgs = existingChain[i].args_hint.split(',').map(a => a.trim()).filter(Boolean);
+        const merged = [...new Set([...oldArgs, ...newArgs])];
+        const mergedHint = merged.join(', ');
+        if (mergedHint !== existingChain[i].args_hint) {
+          existingChain[i].args_hint = mergedHint;
+          updated = true;
+        }
+      }
+      if (updated) {
+        duplicate.tool_chain = JSON.stringify(existingChain);
+        duplicate.success_count++;
+        duplicate.updated_at = Date.now();
+        await bridge.db.saveSkill(duplicate);
+        logInfo(`技能参数优化: ${duplicate.name}`, `合并了新的参数提示`);
+        return 1;
+      }
+    } catch { /* 更新失败不阻塞 */ }
    logInfo(`技能已存在，跳过提取: ${duplicate.name}`);
    return 0;
  }
@@ -113,20 +135,37 @@ export async function extractSkillsFromToolRecords(
 }

 /**
- * 匹配用户消息与已有技能
- * 返回匹配度最高的技能列表
+ * 匹配用户消息与已有技能。
+ * v1.1: 关键词 + 语义匹配融合，未使用技能衰减。
 */
 export async function matchSkills(userMessage: string, limit = 3): Promise<Skill[]> {
  const bridge = window.metonaDesktop;
  if (!bridge?.db?.getAllSkills) return [];

  try {
-    const allSkills: Skill[] = await bridge.db.getAllSkills();
+    const allSkills = await bridge.db.getAllSkills() as any as Skill[];
    if (allSkills.length === 0) return [];

    const messageLower = userMessage.toLowerCase();
    const messageWords = new Set(messageLower.split(/[\s,，。！？、；：""''（）\[\]【】]+/).filter(w => w.length > 1));

+    // ── 语义匹配（如果 embedding 模型可用）──
+    let userEmbedding: number[] | null = null;
+    try {
+      const memEnabled = state.get<boolean>('memoryEnabled', false);
+      const embedModel = state.get<string>('embeddingModel', '');
+      if (memEnabled && embedModel) {
+        const api = state.get<any>(KEYS.API);
+        if (api?.embed) {
+          const result = await api.embed(embedModel, userMessage);
+          userEmbedding = result.embedding || result.embeddings?.[0] || null;
+        }
+      }
+    } catch { /* 语义匹配失败回退关键词 */ }
+
+    const now = Date.now();
+    const THIRTY_DAYS = 30 * 24 * 3600 * 1000;
+
    const scored = allSkills.map(skill => {
      let score = 0;

@@ -134,24 +173,43 @@ export async function matchSkills(userMessage: string, limit = 3): Promise<Skill
      if (skill.trigger_keywords) {
        const kwSet = new Set(skill.trigger_keywords.toLowerCase().split(',').map(k => k.trim()));
        for (const kw of kwSet) {
-          if (kw && messageLower.includes(kw)) score += 0.4;
+          if (kw && messageLower.includes(kw)) score += 0.3;
        }
      }

      // 名称和描述匹配
      const nameWords = skill.name.toLowerCase().split(/[\s→\-]+/);
      for (const w of nameWords) {
-        if (w.length > 1 && messageWords.has(w)) score += 0.2;
+        if (w.length > 1 && messageWords.has(w)) score += 0.15;
      }

-      // 成功率加成
+      // 语义匹配：余弦相似度加权（主导分数）
+      if (userEmbedding && skill.summary) {
+        const skillEmbed = getCachedSkillEmbedding(skill.id);
+        if (skillEmbed) {
+          const sim = cosineSimilarity(userEmbedding, skillEmbed);
+          score += sim * 0.4; // 语义匹配主导
+        }
+      }
+
+      // 成功率加权
      const total = skill.success_count + skill.fail_count;
      if (total > 0) score *= (0.5 + 0.5 * (skill.success_count / total));

+      // ── v1.1 未使用技能衰减 ──
+      if (skill.last_used_at) {
+        const daysSinceUse = (now - skill.last_used_at) / (24 * 3600 * 1000);
+        if (daysSinceUse > 30) score *= 0.5;  // 30天未用降权50%
+        if (daysSinceUse > 90) score *= 0.3;  // 90天未用几乎忽略
+      }
+
      return { skill, score };
    });

-    return scored
+    // ── v1.1 技能链合并：相同前缀的链抽象为一个技能 ──
+    const merged = mergeSimilarSkills(scored);
+
+    return merged
      .filter(s => s.score >= MATCH_THRESHOLD)
      .sort((a, b) => b.score - a.score)
      .slice(0, limit)
@@ -161,6 +219,98 @@ export async function matchSkills(userMessage: string, limit = 3): Promise<Skill
  }
 }

+// ── 向量计算工具 ──
+
+/** 技能 embedding 缓存（LRU 简单实现） */
+const _skillEmbedCache = new Map<string, { embedding: number[]; time: number }>();
+
+function getCachedSkillEmbedding(skillId: string): number[] | null {
+  const cached = _skillEmbedCache.get(skillId);
+  if (cached && Date.now() - cached.time < 300_000) return cached.embedding; // 5分钟有效
+  return null;
+}
+
+export function cacheSkillEmbedding(skillId: string, embedding: number[]): void {
+  _skillEmbedCache.set(skillId, { embedding, time: Date.now() });
+  // 限制缓存大小
+  if (_skillEmbedCache.size > 200) {
+    const oldest = [..._skillEmbedCache.entries()].sort((a, b) => a[1].time - b[1].time)[0];
+    if (oldest) _skillEmbedCache.delete(oldest[0]);
+  }
+}
+
+function cosineSimilarity(a: number[], b: number[]): number {
+  if (!a || !b || a.length !== b.length) return 0;
+  let dot = 0, normA = 0, normB = 0;
+  for (let i = 0; i < a.length; i++) {
+    dot += a[i] * b[i];
+    normA += a[i] * a[i];
+    normB += b[i] * b[i];
+  }
+  const denom = Math.sqrt(normA) * Math.sqrt(normB);
+  return denom === 0 ? 0 : dot / denom;
+}
+
+// ── 技能链合并 ──
+
+interface ScoredSkill { skill: Skill; score: number; }
+
+/**
+ * v1.1 技能链合并：两个以上技能有相同前缀（前 2 步相同），
+ * 合并为一个抽象技能，分数取最高。
+ */
+function mergeSimilarSkills(scored: ScoredSkill[]): ScoredSkill[] {
+  if (scored.length < 2) return scored;
+
+  const result: ScoredSkill[] = [];
+  const used = new Set<number>();
+
+  for (let i = 0; i < scored.length; i++) {
+    if (used.has(i)) continue;
+    const group: ScoredSkill[] = [scored[i]];
+
+    // 找同前缀的
+    for (let j = i + 1; j < scored.length; j++) {
+      if (used.has(j)) continue;
+      if (getChainPrefix(scored[i].skill) === getChainPrefix(scored[j].skill)) {
+        group.push(scored[j]);
+        used.add(j);
+      }
+    }
+
+    if (group.length >= 2) {
+      // 合并为一个抽象技能
+      const best = group.reduce((a, b) => a.score > b.score ? a : b);
+      const chainNames = group.map(g => {
+        try { return (JSON.parse(g.skill.tool_chain) as ToolChainStep[]).map(s => s.name).join('→'); }
+        catch { return g.skill.name; }
+      });
+      const mergedSkill: Skill = {
+        ...best.skill,
+        name: `${getChainPrefix(best.skill)} → ...`,
+        description: `合并了 ${group.length} 个相似技能: ${chainNames.join(' | ')}`,
+        summary: best.skill.summary,
+      };
+      result.push({ skill: mergedSkill, score: best.score * 1.1 }); // 合并后略加分
+      used.add(i);
+    } else {
+      result.push(scored[i]);
+    }
+  }
+
+  return result;
+}
+
+/** 获取技能链的前 2 步前缀（用于相似度判断） */
+function getChainPrefix(skill: Skill): string {
+  try {
+    const chain: ToolChainStep[] = JSON.parse(skill.tool_chain);
+    return chain.slice(0, 2).map(s => s.name).join('→');
+  } catch {
+    return skill.name;
+  }
+}
+
 /**
 * 从匹配的技能构建 system prompt 上下文（Level 0：仅名称 + 描述索引）
 * 完整工具链通过 skill_view 工具按需加载（Level 1）
@@ -286,8 +436,8 @@ export async function listSkills(): Promise<Array<{
 }>> {
  const bridge = window.metonaDesktop;
  if (!bridge?.db?.getAllSkills) return [];
-  const allSkills: Skill[] = await bridge.db.getAllSkills();
-  return allSkills.map(skill => {
+  const allSkills: Skill[] = await bridge.db.getAllSkills() as any;
+  return (allSkills as any[]).map((skill: any) => {
    const total = skill.success_count + skill.fail_count;
    const rate = total > 0 ? Math.round(skill.success_count / total * 100) : 100;
    let chainPreview = '';
@@ -313,7 +463,7 @@ export async function viewSkill(skillName: string): Promise<{
 }> {
  const bridge = window.metonaDesktop;
  if (!bridge?.db?.getAllSkills) return { success: false, error: '数据库不可用' };
-  const allSkills: Skill[] = await bridge.db.getAllSkills();
+  const allSkills: Skill[] = await bridge.db.getAllSkills() as any;
  const skill = allSkills.find(s =>
    s.name.toLowerCase() === skillName.toLowerCase() ||
    s.name.toLowerCase().includes(skillName.toLowerCase())
@@ -6,7 +6,8 @@
 import { state, KEYS } from '../state/state.js';
 import { VectorStore } from './vector-store.js';
 import { logMemory, logDebug, logError } from './log-service.js';
-import type { MemoryEntry, OllamaAPI, VectorItem, SearchResult } from '../types.js';
+import type { MemoryEntry, VectorItem, SearchResult } from '../types.js';
+import type { OllamaAPI } from '../api/ollama.js';

 const MEMORY_COLLECTION_NAME = '记忆向量索引';

@@ -228,7 +228,7 @@ export class VectorStore {

    if (saved && saved.version === this._indexVersion(vectors)) {
      if (!(saved.invertedLists instanceof Map)) {
-        saved.invertedLists = new Map(Object.entries(saved.invertedListsObj || {}));
+        saved.invertedLists = new Map(Object.entries(saved.invertedListsObj || {}).map(([k, v]) => [parseInt(k), v as string[]]));
      }
      this._indexCache.set(colId, saved);
      return saved;
@@ -2,7 +2,9 @@
 * State - 轻量级响应式状态管理
 */

-import type { StateKey, ChatSession, ChatDB, OllamaAPI } from '../types.js';
+import type { StateKey, ChatSession } from '../types.js';
+import type { ChatDB } from '../db/chat-db.js';
+import type { OllamaAPI } from '../api/ollama.js';

 function shallowEqual(a: unknown, b: unknown): boolean {
  if (a === b) return true;
@@ -45,7 +47,7 @@ class AppState {
    }
  }

-  update(key: string, updater: (old: unknown) => unknown): void {
+  update(key: string, updater: (old: any) => any): void {
    const old = this._state[key];
    const value = updater(old);
    this._state[key] = value;
@@ -900,6 +900,75 @@ html, body {
    box-shadow: none;
 }

+/* ── 系统提示词折叠卡片 ── */
+.system-prompt-card {
+    margin-bottom: 10px;
+    border: 1px solid rgba(155, 126, 216, 0.12);
+    border-radius: var(--radius-md);
+    overflow: hidden;
+    background: rgba(155, 126, 216, 0.03);
+}
+
+.system-prompt-header {
+    display: flex;
+    align-items: center;
+    gap: 6px;
+    padding: 6px 12px;
+    cursor: pointer;
+    user-select: none;
+    font-size: 12px;
+    color: var(--text-tertiary);
+    transition: background 0.15s;
+}
+.system-prompt-header:hover {
+    background: rgba(155, 126, 216, 0.06);
+}
+
+.sp-icon { font-size: 14px; flex-shrink: 0; }
+.sp-title { font-weight: 600; color: #9B7ED8; white-space: nowrap; }
+.sp-tokens {
+    font-size: 10px;
+    color: var(--text-tertiary);
+    background: var(--bg-layer);
+    padding: 1px 6px;
+    border-radius: 10px;
+    white-space: nowrap;
+}
+.sp-preview {
+    flex: 1;
+    min-width: 0;
+    overflow: hidden;
+    text-overflow: ellipsis;
+    white-space: nowrap;
+    font-size: 11px;
+    color: var(--text-tertiary);
+}
+.sp-chevron {
+    flex-shrink: 0;
+    transition: transform 0.2s;
+    color: var(--text-tertiary);
+}
+.sp-chevron.sp-expanded { transform: rotate(180deg); }
+
+.system-prompt-body {
+    border-top: 1px solid rgba(155, 126, 216, 0.08);
+    max-height: 300px;
+    overflow-y: auto;
+}
+.system-prompt-body pre {
+    margin: 0;
+    padding: 10px 14px;
+    font-size: 11px;
+    line-height: 1.5;
+    color: var(--text-secondary);
+    white-space: pre-wrap;
+    word-break: break-word;
+    background: none;
+    border: none;
+    box-shadow: none;
+    font-family: var(--font);
+}
+
 /* 打字光标 */
 .typing-cursor {
    display: inline-block;
@@ -51,6 +51,7 @@ export interface OllamaModelInfo {

 export interface OllamaModelDetail {
  capabilities?: string[];
+  model_info?: Record<string, any>;
  [key: string]: unknown;
 }

@@ -59,6 +60,7 @@ export interface ModelCaps {
  vision: boolean;
  tools: boolean;
  completion: boolean;
+  contextLength: number;
 }

 export interface OllamaVersionResponse {
@@ -111,6 +113,8 @@ export interface ChatMessage {
  toolCalls?: ToolCallRecord[];
  /** 标记此消息为 LLM 压缩摘要生成 */
  compressed?: boolean;
+  /** 实际发送给模型的完整系统提示词（调试用） */
+  _systemPrompt?: string;
 }

 export interface ChatSession {