zen-gitsync 2.17.8 → 2.17.10

package/README.md

@@ -166,6 +166,8 @@ In either mode, click **AI Generate** to fill in the fields automatically based
 
 > Click the directory name in the header (or the folder icon) to open this dialog. Type a path, hit **浏览** to use the OS file picker, or pick from **常用目录** for one-click switching. **使用新标签打开** spawns a new GUI tab on that path so you can keep the current project open.
 
+When the GUI is opened on a directory that is not a Git repository, the right pane shows the **Recent projects** list instead — every recent directory with its Git badges (behind / ahead / uncommitted) and one-click "open in a new tab". Each page load runs a `git fetch` pass over all of them automatically, so the ahead/behind badges show the real state rather than the snapshot from the last fetch; the **刷新全部** button does the same thing on demand.
+
 ---
 
 ### Branch Management
@@ -391,6 +393,7 @@ A dedicated view (robot icon in the activity bar) for chatting with the built-in
 | Live session entry | Sending the first message of a new session makes it show up in the list **immediately** with a "Generating..." badge, instead of waiting for the whole turn to finish; once the reply ends and the server persists the session, the entry is replaced by the real timestamp and message count |
 | Streaming chat | SSE-based real-time streaming with thinking process, content, tool calls, and tool results rendered inline |
 | Tool call display | Each tool invocation (run_command, read_file, edit_file, list_files, search_text, write_file) is shown as a collapsible card with arguments preview and execution result |
+| Recent-projects awareness | Ask "which of my projects need a pull?" and the agent calls its built-in `list_projects` tool instead of scanning the disk: it returns exactly the list behind the GUI's **Recent projects** panel (recent directories plus any directory a task was created in, with branch / ahead / behind / uncommitted counts and task progress), so the agent's answer and the UI agree. Ahead/behind reads local refs, so the agent can pass `refresh=true` to run a `git fetch` pass first when the question is about pulling |
 | Session persistence | All conversations are saved to `~/.zen-gitsync/agent-sessions/` as JSON files; the CLI agent (`g ai`) writes to the same directory so Web and CLI sessions are unified |
 | Per-turn tool limit | A single message may trigger up to N tool calls in a row (default **200**, range 1–2000). Configurable in **Settings → AI models → Agent Runtime**; hitting the limit ends the turn and asks you to send another message. The same setting drives the CLI agent |
 | Preset questions | Quick-start buttons on the welcome screen for common tasks (view project structure, analyze code quality, write tests, check git status) |
@@ -779,6 +782,8 @@ $ ZEN_ALLOWED_ORIGINS="https://zen.example.com,http://10.0.0.5:8080" g ui
 
 > 点击顶部条里的目录名（或文件夹图标）即可弹出该对话框。直接输入路径、点击 **浏览** 唤起系统文件选择器，或从 **常用目录** 一键切换。**使用新标签打开** 会在新 GUI 标签里加载目标路径，原项目保持不动。
 
+当 GUI 打开在一个**不是 Git 仓库**的目录上时，右侧会改为显示「最近项目」列表 —— 每个最近目录一张卡片，带 Git 徽标（落后 / 领先 / 未提交）与「在新标签页打开」。每次打开界面时会自动对所有项目跑一遍 `git fetch`，让「领先/落后」显示真实状态而不是上次 fetch 时的快照；**刷新全部** 按钮可以随时手动再刷一遍。
+
 ---
 
 ### 分支管理
@@ -1006,6 +1011,7 @@ Activity Bar 中的机器人图标视图，可直接在浏览器中与内置 AI
 | 会话实时入列 | 新会话发出第一条消息后，左侧列表**立刻**出现这一条（带「正在生成中…」标记），不用等整轮回答跑完；回答结束、服务端落盘后自动替换成真实的时间与条数 |
 | 流式对话 | 基于 SSE 的实时流式输出，包含思考过程、正文内容、工具调用和工具结果的内联渲染 |
 | 工具调用展示 | 每次工具调用（run_command、read_file、edit_file、list_files、search_text、write_file）以可折叠卡片形式展示，含参数预览和执行结果 |
+| 最近项目感知 | 问「我哪些项目需要 pull」时，智能体调用内置的 `list_projects` 工具，而不是自己去扫盘：返回的就是 GUI「最近项目」面板那份清单（最近目录 + 建过任务的目录，带分支 / 领先 / 落后 / 未提交数与任务进度），回答与界面对得上。领先/落后读的是本地引用，因此问到"要不要拉"时它可以带 `refresh=true` 先联网 fetch 一轮再答 |
 | 会话持久化 | 所有对话保存为 JSON 文件到 `~/.zen-gitsync/agent-sessions/`；CLI 智能体（`g ai`）写入同一目录，Web 端与 CLI 端会话统一管理 |
 | 单轮工具调用上限 | 一条消息内智能体最多连续调用多少次工具（默认 **200**，可调范围 1–2000）。在 **设置 → AI 模型配置 → 智能体运行时** 中修改；达到上限本轮会被强制结束并提示再发一条消息继续。CLI 智能体共用同一项设置 |
 | 预设问题 | 开场界面提供快捷按钮（查看项目结构、分析代码质量、写测试、Git 状态检查）|

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "zen-gitsync",
-  "version": "2.17.8",
+  "version": "2.17.10",
   "description": "Auto commit, scheduled sync, and visual GUI for Git. Run `g` in any repo for one-key commit & push, AI commit messages, scheduled background sync, and a drag-and-drop workflow builder.",
   "main": "index.js",
   "type": "module",
@@ -36,6 +36,7 @@
     "verify:wb-proj-branch": "node scripts/verify-workbench-proj-branch.cjs",
     "verify:wb-proj-filter": "node scripts/verify-wb-proj-filter.cjs",
     "verify:wb-panel-resize": "node scripts/verify-wb-panel-resize.cjs",
+    "verify:wb-toolcalls": "node scripts/verify-workbench-toolcalls.mjs",
     "verify:wb-responsive": "node scripts/verify-wb-responsive.cjs",
     "verify:wb-editor-persist": "node scripts/verify-workbench-editor-persist.cjs",
     "e2e": "cd src/ui/client && npx playwright test",

package/src/cli/ai/agent.js

@@ -20,7 +20,7 @@
 //   - 默认使用 g ui 里配置的模型(~/.zen-gitsync/config.json 顶层 models,
 //     isDefault 优先,否则取第一个);--model=<序号|名称> 或会话内 /model 切换
 //   - OpenAI 兼容流式 function calling:模型可以调用 run_command /
-//     read_file / write_file / edit_file / list_files / search_text
+//     read_file / write_file / edit_file / list_files / search_text / ask_user
 //   - 图片:Alt+V 粘贴剪贴板图片、/image <路径> 附加本地图片,
 //     以多模态(image_url)消息发给模型(需模型支持视觉)
 //   - 权限:启动目录内全开放,其他目录可读写;唯一红线由 safety.js 拦截
@@ -44,8 +44,9 @@ import chalk from 'chalk'
 import config from '../../config.js'
 import { getCwd } from '../../utils/index.js'
 import { registerCleanup } from '../cleanup.js'
-import { terminateCommand } from './tools.js'
+import { terminateCommand, TOOL_DEFINITIONS } from './tools.js'
 import { runAgentTurn } from './turn.js'
+import { AgentExtensions } from './extensions.js'
 import { repairToolHistory, loadProjectInstructions } from './context.js'
 import {
   printBanner, printHelpPanel,
@@ -96,6 +97,14 @@ const STRINGS = {
     thinkingHint: '… 已显示前 12 行；输入 /think full 开启后续完整思考',
     emptyResponse: '模型没有返回回答，请重试或切换模型。',
     chars: '字符',
+    extLoadingMcp: (n) => `正在连接 ${n} 个 MCP 服务…`,
+    extLoaded: (skills, tools) => `扩展已就绪: ${skills} 个 Skill · ${tools} 个 MCP 工具(/skills 查看)`,
+    extNone: '',
+    extSkillTitle: '已安装的 Skill',
+    extSkillEmpty: '没有装 Skill。可以在 g ui → 智能体 → Skill 广场 里挑一个装到当前项目或 g ai。',
+    extMcpTitle: '已接入的 MCP 服务',
+    extMcpEmpty: '没有接入 MCP 服务。可以在 g ui → 智能体 → MCP 广场 里挑一个装到当前项目或 g ai。',
+    extMcpFailed: '连接失败',
     imagePasting: '正在读取剪贴板图片…',
     imageAttached: (n, size) => `📎 图片 #${n} 已附加(${size}),将随下一条消息发送`,
     imageEmpty: '剪贴板中没有图片;可先截图再按 Alt+V,或用 /image <路径> 附加本地图片',
@@ -145,6 +154,14 @@ const STRINGS = {
     thinkingHint: '… first 12 lines shown; enter /think full for future reasoning',
     emptyResponse: 'The model returned no answer. Retry or switch models.',
     chars: 'chars',
+    extLoadingMcp: (n) => `Connecting to ${n} MCP server(s)…`,
+    extLoaded: (skills, tools) => `Extensions ready: ${skills} skill(s) · ${tools} MCP tool(s) (/skills for details)`,
+    extNone: '',
+    extSkillTitle: 'Installed skills',
+    extSkillEmpty: 'No skills installed. Browse g ui → Agent → Skill marketplace to install one into this project or g ai.',
+    extMcpTitle: 'Connected MCP servers',
+    extMcpEmpty: 'No MCP servers connected. Browse g ui → Agent → MCP marketplace to install one.',
+    extMcpFailed: 'failed',
     imagePasting: 'Reading clipboard image…',
     imageAttached: (n, size) => `📎 Image #${n} attached (${size}); sent with your next message`,
     imageEmpty: 'No image in clipboard; take a screenshot first, or use /image <path>',
@@ -176,14 +193,28 @@ function modelLabel(m) {
 // ──────────────────────────────────────────────
 // 系统 prompt — 明确告知环境、权限边界与工作方式
 // ──────────────────────────────────────────────
+// 顺序固定:基础人设 → 项目说明(AGENTS.md/CLAUDE.md)→ 扩展能力(skill 清单 + MCP 工具)。
+// 把扩展放最后,是因为它是"增量能力",不该插在权限与环境说明中间打断阅读。
+// state 里存着已加载的 AgentExtensions;没加载(如单测直接调)时 extra 为空,行为与改动前一致。
+function extensionSuffix(state) {
+  try {
+    return state?.extensions?.promptSuffix?.(state.locale) || ''
+  } catch {
+    return ''
+  }
+}
+
 async function buildProjectPrompt(options) {
-  return buildSystemPrompt(options) + await loadProjectInstructions(options.cwd)
+  return buildSystemPrompt(options) + await loadProjectInstructions(options.cwd) + (options.extra || '')
 }
 
 function buildSystemPrompt({ cwd, locale, shellDesc }) {
   const zh = !String(locale || '').startsWith('en')
   const now = new Date().toLocaleString()
   const isWin = process.platform === 'win32'
+  // 工具数量动态取自 tools.js —— 之前这里写死过"6 个",加 ask_user 时没人想起来改,
+  // 就变成了对模型撒谎。数字跟着定义走,以后再加工具不会漏。
+  const builtinToolCount = TOOL_DEFINITIONS.length
   if (zh) {
     return `你是 "g ai" —— zen-gitsync CLI 内置的终端编码智能体,通过工具在用户真实电脑上完成编码任务。
 
@@ -207,6 +238,21 @@ ${isWin ? `- 当前是 Windows cmd.exe,以下 Unix 命令**不存在**,用了必
   · 查命令路径 → cmd 的 where(不是 which)
 - 必须跑 shell 时优先跨平台写法(如 node -e "..."),别用 Unix 专属命令` : `- 当前是 POSIX 环境,Unix 命令可用`}
 
+# 远程仓库(GitHub / Gitee)
+- 问"这个项目关联哪个远端" → run_command 跑 \`git remote -v\`:本地信息,不联网、不依赖任何 CLI,任何环境都能答
+- 问"我账号下有哪些仓库"或某仓库的 PR / Issue → 用官方 CLI。凭据由 CLI 自己保管:
+  · GitHub → \`gh\`。列仓库 \`gh repo list --limit 50 --json name,visibility,updatedAt,primaryLanguage\`;
+    看详情 \`gh repo view <owner/repo>\`;PR \`gh pr list\`;登录态 \`gh auth status\`
+  · Gitee → \`gitee\`。列仓库 \`gitee repo list\`;登录态 \`gitee auth status\`
+    (未登录时退出码**仍是 0**,必须看 \`--json\` 里的 status 字段,只看退出码会误判成已登录)
+  · 两者默认只覆盖**当前登录账号**;组织仓库、私有仓库可能需要更大的 token scope,拉不到就如实说明,
+    不要用其他途径绕过
+  · 绝不向用户索要 token —— 也不要让用户把 token 粘进对话
+- CLI 可能没装、或没在当前进程的 PATH 里(装完没重启服务就是这个表现):
+  报"不是内部或外部命令" / command not found 时,**不要换写法反复重试**,更不要凭印象编仓库名或目录名。
+  直接告诉用户未检测到该 CLI,可在 ZenGitSync 的「远程仓库」页一键安装/登录;
+  若只是想回答本项目的问题,退回 \`git remote -v\`
+
 # 权限(用户已明确授权,无需反复征求同意)
 - 工作目录内:读写文件、执行命令等所有操作直接执行
 - 其他目录:同样可以读取和修改
@@ -226,7 +272,8 @@ ${isWin ? `- 当前是 Windows cmd.exe,以下 Unix 命令**不存在**,用了必
 
 # 与用户交互
 - 需要向用户确认、提问或汇报重要决策时,直接用普通文本输出 —— 用户能实时看到你的文本;
-  不要调用不存在的工具,可用工具只有上面列出的 6 个
+  需要暂停当前任务并等待用户决定或补充信息时,调用 ask_user,不要猜测或只在普通文本里提问
+  不要调用不存在的工具,内置工具就是上面列出的 ${builtinToolCount} 个(若用户装了 MCP,你的工具表里还会多出 mcp__ 开头的工具)
 - 用户可能通过 Alt+V 或 /image 附加图片:图片以 image_url 部件出现在 user 消息里;
   如果当前模型不支持视觉(带图请求报错),提醒用户换用支持视觉的模型
 - 发现高风险或状态不一致的情况(例如版本号 / git tag / CHANGELOG 对不上、发布前环境异常、
@@ -258,6 +305,24 @@ ${isWin ? `- This is Windows cmd.exe. The following Unix commands do NOT exist h
   · Find executable → cmd's where (not which)
 - For shell one-offs, prefer cross-platform forms like node -e "..."` : `- POSIX environment: Unix commands are available`}
 
+# Remote repositories (GitHub / Gitee)
+- "Which remote does this project point at?" → run_command \`git remote -v\`: local info, no network, no CLI needed
+- "Which repos do I have?" or a repo's PRs / issues → use the official CLI. The CLI owns the credentials:
+  · GitHub → \`gh\`. List repos \`gh repo list --limit 50 --json name,visibility,updatedAt,primaryLanguage\`;
+    details \`gh repo view <owner/repo>\`; PRs \`gh pr list\`; auth state \`gh auth status\`
+  · Gitee → \`gitee\`. List repos \`gitee repo list\`; auth state \`gitee auth status\`
+    (the exit code is still 0 when logged out — read the status field from \`--json\`, or you will
+    wrongly report "logged in")
+  · Both default to the **currently authenticated account only**; org and private repos may need
+    broader token scopes. If you cannot see them, say so plainly — do not route around it
+  · Never ask the user for a token, and never have them paste one into the conversation
+- The CLI may not be installed, or missing from this process's PATH (that is what "installed but the
+  server cannot find it" looks like):
+  on "not recognized" / command not found, do NOT retry with a different spelling and do NOT invent
+  repo or directory names. Tell the user the CLI was not found and point them to ZenGitSync's
+  "Remote repositories" page (one-click install / login); if you only need to answer for this project,
+  fall back to \`git remote -v\`
+
 # Permissions (explicitly granted by the user — do not keep asking)
 - Inside the working directory: read/write files and run commands directly
 - Other directories: may also be read and modified
@@ -265,6 +330,7 @@ ${isWin ? `- This is Windows cmd.exe. The following Unix commands do NOT exist h
 
 # How to work
 - Act first, ask later: use tools (list_files / read_file / search_text / run_command) instead of asking the user
+- When a decision or missing detail must come from the user, call ask_user. Do not guess or merely describe a question in plain text.
 - After editing code, verify: run tests, builds, or at least a syntax check via run_command
 - Prefer edit_file for precise replacements; read_file first, old_string must match the file exactly
 - Read large files in segments (offset/limit)
@@ -275,7 +341,7 @@ ${isWin ? `- This is Windows cmd.exe. The following Unix commands do NOT exist h
 - For a brief overview, inspect only the necessary structure and entry points; keep exploration proportional to the request
 
 # Talking to the user
-- When you need to confirm something, ask a question, or report an important decision, just write plain text — the user sees your output in real time. Never call tools that do not exist; only the 6 tools listed above are available
+- When you need to confirm something, ask a question, or report an important decision, just write plain text — the user sees your output in real time. Use ask_user when the task must pause for an answer. Never call tools that do not exist; the built-in set is the ${builtinToolCount} tools listed above (if the user installed MCP servers, extra tools prefixed with mcp__ will also appear in your tool list)
 - The user can attach images (Alt+V or /image); they arrive as image_url parts in your user messages. If the current model rejects images (no vision support), tell the user to switch to a vision-capable model
 - When you spot high-risk or inconsistent state (version number / git tag / CHANGELOG mismatch, abnormal release environment, unexpected repo state), explain the finding and its impact in text, then STOP and wait for the user's decision instead of proceeding with destructive operations
 
@@ -285,54 +351,10 @@ ${isWin ? `- This is Windows cmd.exe. The following Unix commands do NOT exist h
 }
 
 // ──────────────────────────────────────────────
-// 消息兼容处理。只作用于请求副本，保留完整会话记录。
+// 消息准备(消毒 / 旧图片降级)统一实现在 context.js —— CLI 与 Web 智能体面板
+// 共用同一份,避免两侧各写一遍后行为分叉。这里 re-export 保持既有引用与测试不断。
 // ──────────────────────────────────────────────
-export function sanitizeMessages(messages) {
-  for (const m of messages) {
-    if (m == null || typeof m !== 'object') continue
-    if (m.content === null || m.content === undefined) {
-      if (m.role === 'assistant') m.content = null
-      continue
-    }
-    if (typeof m.content !== 'string') continue
-    const trimmed = m.content.trim()
-    if (trimmed === '') {
-      if (m.role === 'assistant') {
-        m.content = null
-      } else if (m.role === 'tool') {
-        m.content = '(no output)'
-      } else if (m.role === 'user') {
-        m.content = ' '
-      }
-      continue
-    }
-    if (m.role === 'assistant' && Array.isArray(m.tool_calls) && m.tool_calls.length > 0) {
-      m.content = null
-    }
-  }
-  return messages
-}
-
-// ──────────────────────────────────────────────
-// 多模态历史:base64 图片很占上下文,只保留"最近一条带图消息"里的图片,
-// 更早消息里的 image_url 部件降级为文字占位(模型仍知道这里曾有图)
-// ──────────────────────────────────────────────
-export function stripStaleImages(messages, locale) {
-  const placeholder = String(locale || '').startsWith('en')
-    ? '[image omitted from history]'
-    : '[图片已从历史中省略]'
-  let seenLatest = false
-  for (let i = messages.length - 1; i >= 0; i--) {
-    const m = messages[i]
-    if (m?.role !== 'user' || !Array.isArray(m.content)) continue
-    const hasImage = m.content.some(p => p?.type === 'image_url')
-    if (!hasImage) continue
-    if (!seenLatest) { seenLatest = true; continue }
-    m.content = m.content.map(p => p?.type === 'image_url'
-      ? { type: 'text', text: placeholder }
-      : p)
-  }
-}
+export { sanitizeMessages, stripStaleImages } from './context.js'
 
 // ──────────────────────────────────────────────
 // 单轮 agent 循环:用户一句话 → 流式输出 → 工具调用 → 再调用模型 … 直到模型给出最终文本
@@ -350,6 +372,7 @@ function printSlashHelp(t, locale) {
     '  /think compact    预览前 12 行思考',
     '  /think off        隐藏思考；/think 切换完整与隐藏',
     '  /tools [模式]     compact 精简 / full 完整工具输出',
+    '  /skills           查看已安装的 Skill 与已接入的 MCP 服务',
     '  /stats            查看上一轮与会话累计用量、耗时',
     '  /new              开启新对话',
     '  /resume           选择并恢复之前的对话',
@@ -370,6 +393,7 @@ function printSlashHelp(t, locale) {
     '  /think compact    Preview the first 12 thinking lines',
     '  /think off        Hide thinking; /think toggles full / hidden',
     '  /tools [mode]     compact / full tool output',
+    '  /skills           List installed skills and connected MCP servers',
     '  /stats            Last turn timing and session token usage',
     '  /new              Start a new conversation',
     '  /resume           Choose and resume a previous conversation',
@@ -388,7 +412,7 @@ export async function handleSlashCommand(state, input, t) {
   const [cmd, ...rest] = input.split(/\s+/)
   const arg = rest.join(' ').trim()
   const zh = !String(state.locale).startsWith('en')
-  if (state.busy && !['/help', '/stats', '/think', '/tools', '/exit', '/quit'].includes(cmd)) {
+  if (state.busy && !['/help', '/stats', '/think', '/tools', '/skills', '/mcp', '/exit', '/quit'].includes(cmd)) {
     printDim(t.busy)
     return 'ok'
   }
@@ -467,7 +491,7 @@ export async function handleSlashCommand(state, input, t) {
       state.sessionStats = session.sessionStats || null
       const systemMessage = {
         role: 'system',
-        content: await buildProjectPrompt({ cwd: state.ctx.cwd, locale: state.locale, shellDesc: state.shellDesc }),
+        content: await buildProjectPrompt({ cwd: state.ctx.cwd, locale: state.locale, shellDesc: state.shellDesc, extra: extensionSuffix(state) }),
       }
       if (state.messages[0]?.role === 'system') state.messages[0] = systemMessage
       else state.messages.unshift(systemMessage)
@@ -499,6 +523,26 @@ export async function handleSlashCommand(state, input, t) {
     printOk(zh ? `后续工具结果：${state.fullTools ? '完整' : '精简'}` : `Future tool output: ${state.fullTools ? 'full' : 'compact'}`)
     return 'ok'
   }
+  if (cmd === '/skills' || cmd === '/mcp') {
+    const info = state.extensions?.summary(state.locale) || { skills: [], servers: [] }
+    console.log(chalk.bold(`\n  ${t.extSkillTitle}`))
+    if (info.skills.length === 0) printDim(t.extSkillEmpty)
+    for (const skill of info.skills) {
+      const scope = skill.scope === 'project' ? (zh ? '项目' : 'project') : (zh ? '全局' : 'global')
+      console.log(`  · ${chalk.cyan(skill.name)} ${chalk.dim(`[${scope}]`)}${skill.description ? chalk.dim(` — ${skill.description}`) : ''}`)
+    }
+    console.log(chalk.bold(`\n  ${t.extMcpTitle}`))
+    if (info.servers.length === 0) printDim(t.extMcpEmpty)
+    for (const server of info.servers) {
+      if (server.error) {
+        console.log(`  · ${chalk.red(server.id)} ${chalk.dim(`— ${t.extMcpFailed}: ${server.error}`)}`)
+      } else {
+        console.log(`  · ${chalk.cyan(server.id)} ${chalk.dim(`— ${server.tools} ${zh ? '个工具' : 'tools'} (${server.command})`)}`)
+      }
+    }
+    console.log('')
+    return 'ok'
+  }
   if (cmd === '/stats') {
     if (state.lastTurnStats) printTurnSummary(state.lastTurnStats, { locale: state.locale, session: state.sessionStats })
     else printDim(zh ? '完成一轮对话后显示用量与耗时。' : 'Usage and timing appear after a turn completes.')
@@ -602,8 +646,10 @@ export async function handleSlashCommand(state, input, t) {
       const st = await fsp.stat(target)
       if (!st.isDirectory()) throw new Error('not a dir')
       state.ctx.cwd = target
-      // 同步更新 system prompt 里的 cwd 说明(重建首条消息)
-      state.messages[0] = { role: 'system', content: await buildProjectPrompt({ cwd: target, locale: state.locale, shellDesc: state.shellDesc }) }
+      // 项目级 skill 与 .mcp.json 跟着工作目录走:先重载扩展(会关掉旧目录起的 MCP 子进程),
+      // 再重建首条 system 消息 —— 顺序反了会把旧目录的 skill 清单写进新目录的提示词。
+      await state.extensions?.reload({ cwd: target, locale: state.locale, onWarn: printWarn })
+      state.messages[0] = { role: 'system', content: await buildProjectPrompt({ cwd: target, locale: state.locale, shellDesc: state.shellDesc, extra: extensionSuffix(state) }) }
       await state.persistSession?.()
       printOk(t.cdOk(target))
     } catch {
@@ -660,12 +706,21 @@ export async function runAiAgent(argv = []) {
     ? 'cmd.exe (Windows CMD)'
     : (process.env.SHELL || '/bin/sh')
 
+  // 加载扩展(Skill 清单 + MCP 服务)。MCP 要起子进程,npx 首次还得下载包,
+  // 所以这一步可能明显停顿 —— 有 MCP 时先提示一句。任何失败都只降级并打印原因。
+  const extensions = await AgentExtensions.load({
+    cwd,
+    locale,
+    onWarn: printWarn,
+    onNotice: (count) => printDim(t.extLoadingMcp(count)),
+  })
+
   const state = {
     // 会话持久化:CLI 对话也保存到 ~/.zen-gitsync/agent-sessions/,
     // 在 Web UI 智能体 tab 中可见(带 CLI 标记)
     sessionId: genSessionId(),
     sessionCreatedAt: new Date().toISOString(),
-    messages: [{ role: 'system', content: await buildProjectPrompt({ cwd, locale, shellDesc }) }],
+    messages: [{ role: 'system', content: await buildProjectPrompt({ cwd, locale, shellDesc, extra: extensions.promptSuffix(locale) }) }],
     ctx: {
       cwd,
       onChild: (child) => {
@@ -690,11 +745,11 @@ export async function runAiAgent(argv = []) {
     fullTools: false,
     lastTurnStats: null,
     sessionStats: null,
-    prepareMessages: (messages) => { stripStaleImages(messages, locale); sanitizeMessages(messages) },
     pendingImages: [],      // Alt+V / /image 附加的待发送图片 [{path, bytes}]
     pasting: false,         // 剪贴板读取进行中(防 Alt+V 连打并发)
     inWizard: false,        // /addmodel 交互式向导进行中:忽略 REPL 的 line 事件
     rl: null,               // REPL readline 引用(供 /addmodel 复用做 rl.question)
+    extensions,             // 已加载的 Skill + MCP 扩展(见 extensions.js)
   }
 
   // 持久化当前会话到磁盘(供 Web UI 读取)
@@ -733,13 +788,21 @@ export async function runAiAgent(argv = []) {
     state.cancelRequested = true
     try { state.abortController?.abort() } catch (_) {}
     await terminateCommand(state.currentChild)
+    // MCP 子进程必须显式收掉:它们挂在 stdio 上,不收会让进程无法自然退出
+    await extensions.close()
     await persistSession()
   })
 
+  // 扩展摘要:一条 dim 行,没有装扩展时完全静默(extNone 为空串)
+  const extSummary = extensions.skills.length || extensions.tools.length
+    ? t.extLoaded(extensions.skills.length, extensions.tools.length)
+    : t.extNone
+
   // 单发模式:g ai "帮我改个 bug"(位置参数拼成 prompt,跑一轮就退出)
   const oneShot = argv.filter(a => !a.startsWith('-')).join(' ').trim()
   if (oneShot) {
     console.log(chalk.dim(`[${t.bannerModel}] ${modelLabel(model)} · [${t.bannerCwd}] ${cwd}`))
+    if (extSummary) printDim(extSummary)
     state.abortController = new AbortController()
     const stats = await runAgentTurn(state, oneShot, t)
     if (stats.status === 'completed') printDim(t.oneShotDone)
@@ -757,6 +820,7 @@ export async function runAiAgent(argv = []) {
     cwdText: t.bannerCwd,
     tip: t.bannerTip,
   })
+  if (extSummary) printDim(extSummary)
 
   const rl = readline.createInterface({
     input: process.stdin,
@@ -767,6 +831,94 @@ export async function runAiAgent(argv = []) {
   })
   state.rl = rl  // 供 /addmodel 等需要 rl.question 的斜杠命令复用
 
+  // ask_user pauses the tool loop inside the same readline session. The REPL
+  // line handler is bypassed while this small wizard owns the input line.
+  state.ctx.askUser = ({ question, options = [], allowFreeText = true }) => {
+    const zh = !String(state.locale || '').startsWith('en')
+    const safeOptions = Array.isArray(options) ? options.filter(Boolean).slice(0, 20) : []
+    state.inWizard = true
+    if (safeOptions.length > 0) {
+      console.log(chalk.cyan(`\n${question}`))
+      safeOptions.forEach((option, index) => console.log(`  ${chalk.cyan(String(index + 1))}. ${option}`))
+    } else {
+      console.log(chalk.cyan(`\n${question}`))
+    }
+
+    return new Promise(resolve => {
+      let settled = false
+      const questionController = new AbortController()
+      const parentSignal = state.abortController?.signal
+      const cleanup = () => {
+        rl.removeListener?.('close', onClose)
+        rl.removeListener?.('operationCancel', onCancel)
+        rl.input?.removeListener?.('keypress', onKeypressCancel)
+        parentSignal?.removeEventListener?.('abort', onParentAbort)
+      }
+      const finish = answer => {
+        if (settled) return
+        settled = true
+        cleanup()
+        state.inWizard = false
+        resolve(answer)
+      }
+      const cancel = () => {
+        if (settled) return
+        questionController.abort()
+        finish('Cancelled by user; the question was not answered.')
+      }
+      const onClose = () => cancel()
+      const onCancel = () => {
+        state.cancelRequested = true
+        try { state.abortController?.abort() } catch (_) {}
+        cancel()
+      }
+      const onParentAbort = () => cancel()
+      const onKeypressCancel = (_str, key) => {
+        if (!key?.ctrl || key.name !== 'c') return
+        key.name = 'escape'
+        key.sequence = ''
+        key.ctrl = false
+        key.meta = false
+        key.shift = false
+        onCancel()
+      }
+
+      rl.once('close', onClose)
+      rl.once('operationCancel', onCancel)
+      rl.input?.prependListener?.('keypress', onKeypressCancel)
+      parentSignal?.addEventListener?.('abort', onParentAbort, { once: true })
+
+      const promptAnswer = () => {
+        const hint = safeOptions.length > 0
+          ? (allowFreeText ? (zh ? '请输入序号或直接输入回答' : 'Choose a number or type an answer') : (zh ? '请输入序号' : 'Choose a number'))
+          : (zh ? '请输入回答' : 'Type your answer')
+        rl.question(`${hint}: `, { signal: questionController.signal }, answer => {
+          if (settled) return
+          const trimmed = String(answer || '').trim()
+          if (safeOptions.length > 0) {
+            const index = Number.parseInt(trimmed, 10)
+            if (Number.isInteger(index) && index >= 1 && index <= safeOptions.length) {
+              finish(safeOptions[index - 1])
+              return
+            }
+            if (!allowFreeText || !trimmed) {
+              console.log(chalk.yellow(zh ? `请输入 1-${safeOptions.length} 之间的序号` : `Enter a number from 1 to ${safeOptions.length}`))
+              promptAnswer()
+              return
+            }
+          }
+          if (!trimmed) {
+            console.log(chalk.yellow(zh ? '回答不能为空' : 'Answer cannot be empty'))
+            promptAnswer()
+            return
+          }
+          finish(trimmed)
+        })
+      }
+      promptAnswer()
+    })
+  }
+
   // 输入提示符管理。
   //
   // readline 的 _refreshLine() 用 prevRows 追踪“提示符占了几行”,

package/src/cli/ai/context.js

@@ -64,6 +64,64 @@ export function buildRequestMessages(messages, { maxMessages = 40, maxChars = 80
   return result
 }
 
+// Provider compatibility. Some providers (Moonshot/Kimi, Zhipu, Volcengine, MiniMax…)
+// reject an assistant message whose content is empty while tool_calls are present
+// ("chat content is empty (2013)"), and implementations disagree on which shapes are
+// legal. Runs on the request copy only — what the model actually produced stays on disk.
+//   - assistant with tool_calls → content forced to null
+//   - assistant with blank content → null
+//   - user with blank content → a single space (null is rejected by some providers)
+//   - tool with blank content → '(no output)' (otherwise it can be dropped while serializing)
+export function sanitizeMessages(messages) {
+  for (const m of messages) {
+    if (m == null || typeof m !== 'object') continue
+    // Non-string content (multimodal user parts, already null) is left alone.
+    if (m.content === null || m.content === undefined) {
+      if (m.role === 'assistant') m.content = null
+      continue
+    }
+    if (typeof m.content !== 'string') continue
+    if (m.content.trim() === '') {
+      if (m.role === 'assistant') m.content = null
+      else if (m.role === 'tool') m.content = '(no output)'
+      else if (m.role === 'user') m.content = ' '
+      continue
+    }
+    if (m.role === 'assistant' && Array.isArray(m.tool_calls) && m.tool_calls.length > 0) {
+      m.content = null
+    }
+  }
+  return messages
+}
+
+// Base64 images dominate the payload, so only the newest image-bearing user message
+// keeps its images; older ones degrade to a text placeholder (the model still knows
+// an image was there). Reassigns `content` on the request copy, never on the transcript.
+export function stripStaleImages(messages, locale) {
+  const placeholder = String(locale || '').startsWith('en')
+    ? '[image omitted from history]'
+    : '[图片已从历史中省略]'
+  let seenLatest = false
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const m = messages[i]
+    if (m?.role !== 'user' || !Array.isArray(m.content)) continue
+    if (!m.content.some(p => p?.type === 'image_url')) continue
+    if (!seenLatest) { seenLatest = true; continue }
+    m.content = m.content.map(p => p?.type === 'image_url' ? { type: 'text', text: placeholder } : p)
+  }
+  return messages
+}
+
+// Single entry point for every outgoing payload. The CLI (`turn.js`) and the GUI agent
+// panel (`agentChat.js`) must both call this instead of assembling their own copy —
+// that is the only thing keeping the two from drifting apart again.
+// Returns a fresh array; the caller's transcript is never modified.
+export function prepareRequestMessages(messages, { locale, maxMessages = 40, maxChars = 80000 } = {}) {
+  const copy = buildRequestMessages(messages, { maxMessages, maxChars })
+  stripStaleImages(copy, locale)
+  return sanitizeMessages(copy)
+}
+
 // Only load explicitly named project instruction files, with bounded local redirects.
 export async function loadProjectInstructions(cwd) {
   const visited = new Set(), sections = []