Expose remote desktop over MCP and document recent UX changes

MCP tool surface
- New ``remote_desktop_tools()`` factory in
  ``je_auto_control/utils/mcp_server/tools/_factories.py`` exposes
  the same singleton remote-desktop registry the GUI's Remote
  Desktop tab uses:
    ac_remote_host_start         (token, bind, port, fps, quality, …)
    ac_remote_host_stop          (timeout)
    ac_remote_host_status        (read-only)
    ac_remote_viewer_connect     (host, port, token, expected_host_id)
    ac_remote_viewer_disconnect  (timeout)
    ac_remote_viewer_status      (read-only)
    ac_remote_viewer_send_input  (action: dict)
- Adapter handlers in ``_handlers.py`` import the registry lazily so
  the existing tool group stays cheap to load.
- Status / observer tools (`*_status`) carry ``READ_ONLY`` so they
  survive ``--readonly`` mode; ``send_input`` is correctly tagged
  ``destructiveHint`` so MCP clients can prompt for confirmation.

Tests
- ``test_mcp_server.test_remote_desktop_tools_are_registered`` —
  schema + annotation sanity check.
- ``test_mcp_server.test_remote_desktop_status_tools_survive_read_only_mode``
  — confirms the read-only filter keeps status tools and drops the
  destructive ones.
- 591 / 591 headless pytest pass; ruff clean.

Docs
- ``docs/source/Eng|Zh/doc/new_features/new_features_doc.rst`` —
  three new sections: AnyDesk-style popout viewer window,
  responsive ``QScrollArea`` sub-tab sizing, and the new
  ``ac_remote_*`` MCP tool surface (with a worked example).
- ``docs/source/Eng|Zh/doc/mcp_server/mcp_server_doc.rst`` — the
  Remote Desktop tool group is listed in the tool catalogue.
- ``README.md`` and the two CN/TW READMEs — Remote Desktop entry
  now mentions the popout, ``QScrollArea`` resizing, and the
  headless / MCP driveability; MCP entry highlights the new
  ``ac_remote_*`` tools and the bumped tool count (~100).

Author: JE-Chen

README.md

@@ -63,7 +63,7 @@
 - **OCR** — extract text from screen regions using Tesseract; wait for, click, or locate rendered text; regex search and full-region dump
 - **LLM Action Planner** — translate a plain-language description into a validated `AC_*` action list using Claude
 - **Runtime Variables & Control Flow** — `${var}` substitution at execution time, plus `AC_set_var` / `AC_inc_var` / `AC_if_var` / `AC_for_each` / `AC_loop` / `AC_retry` for data-driven scripts
-- **Remote Desktop** — stream this machine's screen and accept remote input over a token-authenticated TCP protocol, *or* connect to another machine and view + control it (host + viewer GUIs included). Optional TLS (HTTPS-grade encryption), WebSocket transport (ws:// + wss:// for browser / firewall-friendly clients), persistent 9-digit Host ID, host→viewer audio streaming, bidirectional clipboard sync (text + image), and chunked file transfer (drag-drop + progress bar; arbitrary destination path; no size cap). Plus folder sync (additive mirror — local deletions never propagate) and a self-hosted coturn TURN config bundle generator (turnserver.conf + systemd unit + docker-compose + README)
+- **Remote Desktop** — stream this machine's screen and accept remote input over a token-authenticated TCP protocol, *or* connect to another machine and view + control it (host + viewer GUIs included). Optional TLS (HTTPS-grade encryption), WebSocket transport (ws:// + wss:// for browser / firewall-friendly clients), persistent 9-digit Host ID, host→viewer audio streaming, bidirectional clipboard sync (text + image), and chunked file transfer (drag-drop + progress bar; arbitrary destination path; no size cap). Plus folder sync (additive mirror — local deletions never propagate) and a self-hosted coturn TURN config bundle generator (turnserver.conf + systemd unit + docker-compose + README). **AnyDesk-style popout**: when the viewer authenticates, the live remote desktop opens in its own resizable top-level window so the control panel stays uncluttered. The Remote Desktop tabs are wrapped in `QScrollArea` so the panel stays usable on small windows and stretches edge-to-edge on 4K displays. Driveable headlessly via `je_auto_control` and over MCP through the new `ac_remote_*` tools
 - **Clipboard** — read/write system clipboard text on Windows, macOS, and Linux
 - **Screenshot & Screen Recording** — capture full screen or regions as images, record screen to video (AVI/MP4)
 - **Action Recording & Playback** — record mouse/keyboard events and replay them
@@ -73,7 +73,7 @@
 - **Event Triggers** — fire scripts when an image appears, a window opens, a pixel changes, or a file is modified
 - **Run History** — SQLite-backed run log across scheduler / triggers / hotkeys / REST with auto error-screenshot artifacts
 - **Report Generation** — export test records as HTML, JSON, or XML reports with success/failure status
-- **MCP Server** — JSON-RPC 2.0 Model Context Protocol server (stdio + HTTP/SSE) so Claude Desktop / Claude Code / custom tool-use loops can drive AutoControl. ~90 tools, full protocol coverage (resources, prompts, sampling, roots, logging, progress, cancellation, elicitation), bearer-token auth + TLS, audit log, rate limit, plugin hot-reload, CI fake backend
+- **MCP Server** — JSON-RPC 2.0 Model Context Protocol server (stdio + HTTP/SSE) so Claude Desktop / Claude Code / custom tool-use loops can drive AutoControl. ~100 tools, full protocol coverage (resources, prompts, sampling, roots, logging, progress, cancellation, elicitation), bearer-token auth + TLS, audit log, rate limit, plugin hot-reload, CI fake backend. New in this release: `ac_remote_host_start` / `ac_remote_host_stop` / `ac_remote_host_status` / `ac_remote_viewer_connect` / `ac_remote_viewer_disconnect` / `ac_remote_viewer_status` / `ac_remote_viewer_send_input` wrap the same singleton remote-desktop registry the GUI uses, so a model can spin up a host, open a viewer to another machine, and forward mouse / keyboard / type / hotkey actions through the active session
 - **Remote Automation** — TCP socket server **and** hardened REST API: bearer-token auth, per-IP rate limit + lockout, SQLite audit hook, Prometheus `/metrics`, OpenAPI-style endpoint table (`/health`, `/screen_size`, `/sessions`, `/screenshot`, `/execute`, `/audit/list`, `/audit/verify`, `/inspector/recent`, `/usb/devices`, `/diagnose`, ...), and a vanilla-JS browser dashboard at `/dashboard` (any phone with HTTP reach can monitor the host)
 - **Plugin Loader** — drop `.py` files exposing `AC_*` callables into a directory and register them as executor commands at runtime
 - **Shell Integration** — execute shell commands within automation workflows with async output capture

README/README_zh-CN.md

@@ -62,7 +62,7 @@
 - **OCR** — 使用 Tesseract 从屏幕提取文字，可搜索、点击或等待文字出现；支持 regex 搜索与整块区域 dump
 - **LLM 动作规划器** — 用 Claude 把自然语言描述翻译成验证过的 `AC_*` 动作清单
 - **运行期变量与流程控制** — 执行时 `${var}` 替换，加上 `AC_set_var` / `AC_inc_var` / `AC_if_var` / `AC_for_each` / `AC_loop` / `AC_retry` 让脚本数据驱动
-- **远程桌面** — 用 token 认证的 TCP 协议串流本机画面并接收输入，**或** 连接到他机观看与控制（host + viewer GUI 内置）。可选 TLS（HTTPS 级加密）、WebSocket 传输（``ws://`` + ``wss://``，穿墙／浏览器友好）、持久化 9 位数 Host ID、host→viewer 音频串流、双向剪贴板同步（文字 + 图片）、分块文件传输（拖放 + 进度条；任意目的路径；无大小上限）。另含文件夹同步（增量镜像 — 本地删除不会传出去）与自建 coturn TURN 配置包生成器（turnserver.conf + systemd unit + docker-compose + README）
+- **远程桌面** — 用 token 认证的 TCP 协议串流本机画面并接收输入，**或** 连接到他机观看与控制（host + viewer GUI 内置）。可选 TLS（HTTPS 级加密）、WebSocket 传输（``ws://`` + ``wss://``，穿墙／浏览器友好）、持久化 9 位数 Host ID、host→viewer 音频串流、双向剪贴板同步（文字 + 图片）、分块文件传输（拖放 + 进度条；任意目的路径；无大小上限）。另含文件夹同步（增量镜像 — 本地删除不会传出去）与自建 coturn TURN 配置包生成器（turnserver.conf + systemd unit + docker-compose + README）。**AnyDesk 风格弹出窗口**：viewer 认证成功后远程桌面会开在独立的可调整大小顶层窗口，控制面板保持简洁；Remote Desktop 子分页外层包了 `QScrollArea`，小窗口下可滚动、4K 屏幕下会铺满。同时支持 headless API 与 MCP 工具 (`ac_remote_*`) 直接驱动
 - **剪贴板** — 于 Windows / macOS / Linux 读写系统剪贴板文本
 - **截图与屏幕录制** — 捕获全屏或指定区域为图片，录制屏幕为视频（AVI/MP4）
 - **动作录制与回放** — 录制鼠标/键盘事件并重新播放
@@ -72,7 +72,7 @@
 - **事件触发器** — 检测到图像出现、窗口出现、像素变化或文件变动时自动执行脚本
 - **执行历史** — 使用 SQLite 记录 scheduler / triggers / hotkeys / REST 的执行结果；错误时自动附带截图
 - **报告生成** — 将测试记录导出为 HTML、JSON 或 XML 报告，包含成功/失败状态
-- **MCP 服务器** — JSON-RPC 2.0 Model Context Protocol 服务（stdio + HTTP/SSE），让 Claude Desktop / Claude Code / 自定义 tool-use 循环直接驱动 AutoControl。约 90 个工具,完整协议支持(resources、prompts、sampling、roots、logging、progress、cancellation、elicitation),Bearer token 验证 + TLS、审计 log、rate limit、plugin 热加载、CI fake backend
+- **MCP 服务器** — JSON-RPC 2.0 Model Context Protocol 服务（stdio + HTTP/SSE），让 Claude Desktop / Claude Code / 自定义 tool-use 循环直接驱动 AutoControl。约 100 个工具,完整协议支持(resources、prompts、sampling、roots、logging、progress、cancellation、elicitation),Bearer token 验证 + TLS、审计 log、rate limit、plugin 热加载、CI fake backend。**本次新增** `ac_remote_host_start` / `ac_remote_host_stop` / `ac_remote_host_status` / `ac_remote_viewer_connect` / `ac_remote_viewer_disconnect` / `ac_remote_viewer_status` / `ac_remote_viewer_send_input` 包装 GUI 远程桌面分页所用的 process-global registry，模型可以直接启动 host、连线 viewer、转发 mouse／keyboard／type／hotkey 动作
 - **远程自动化** — TCP Socket 服务器 **加上** 强化版 REST API：bearer token 认证、per-IP 速率限制 + 失败锁定、SQLite 审计 hook、Prometheus `/metrics`、完整端点列表（`/health`、`/screen_size`、`/sessions`、`/screenshot`、`/execute`、`/audit/list`、`/audit/verify`、`/inspector/recent`、`/usb/devices`、`/diagnose`、…），以及 vanilla-JS 的浏览器 dashboard `/dashboard`（任何能 HTTP 连到主机的手机都能监控）
 - **插件加载器** — 将定义 `AC_*` 可调用对象的 `.py` 文件放入目录，运行时即可注册为 executor 命令
 - **Shell 集成** — 在自动化流程中执行 Shell 命令，支持异步输出捕获

README/README_zh-TW.md

@@ -62,7 +62,7 @@
 - **OCR** — 使用 Tesseract 從螢幕擷取文字，可搜尋、點擊或等待文字出現；支援 regex 搜尋與整塊區域 dump
 - **LLM 動作規劃器** — 用 Claude 把自然語言描述翻譯成驗證過的 `AC_*` 動作清單
 - **執行期變數與流程控制** — 執行時 `${var}` 取代，加上 `AC_set_var` / `AC_inc_var` / `AC_if_var` / `AC_for_each` / `AC_loop` / `AC_retry` 讓腳本資料驅動
-- **遠端桌面** — 用 token 認證的 TCP 協定串流本機畫面並接收輸入，**或** 連線到他機觀看與控制（host + viewer GUI 皆內建）。可選 TLS（HTTPS 級加密）、WebSocket 傳輸（``ws://`` + ``wss://``，穿牆／瀏覽器友善）、持久化 9 位數 Host ID、host→viewer 音訊串流、雙向剪貼簿同步（文字 + 圖片）、分塊檔案傳輸（拖放 + 進度條；任意目的路徑；無大小上限）。另含資料夾同步（增量鏡像 — 本地刪除不會傳出去）與自架 coturn TURN 設定包產生器（turnserver.conf + systemd unit + docker-compose + README）
+- **遠端桌面** — 用 token 認證的 TCP 協定串流本機畫面並接收輸入，**或** 連線到他機觀看與控制（host + viewer GUI 皆內建）。可選 TLS（HTTPS 級加密）、WebSocket 傳輸（``ws://`` + ``wss://``，穿牆／瀏覽器友善）、持久化 9 位數 Host ID、host→viewer 音訊串流、雙向剪貼簿同步（文字 + 圖片）、分塊檔案傳輸（拖放 + 進度條；任意目的路徑；無大小上限）。另含資料夾同步（增量鏡像 — 本地刪除不會傳出去）與自架 coturn TURN 設定包產生器（turnserver.conf + systemd unit + docker-compose + README）。**AnyDesk 風格彈出視窗**：viewer 認證成功後遠端桌面會開在獨立的可調整大小頂層視窗，控制面板維持簡潔；Remote Desktop 子分頁外層包了 `QScrollArea`，小視窗下可捲動、4K 螢幕下會延展到整寬。同時可由 headless API 與 MCP 工具（`ac_remote_*`）直接驅動
 - **剪貼簿** — 於 Windows / macOS / Linux 讀寫系統剪貼簿文字
 - **截圖與螢幕錄製** — 擷取全螢幕或指定區域為圖片，錄製螢幕為影片（AVI/MP4）
 - **動作錄製與回放** — 錄製滑鼠/鍵盤事件並重新播放
@@ -72,7 +72,7 @@
 - **事件觸發器** — 偵測到影像出現、視窗出現、像素變化或檔案變動時自動執行腳本
 - **執行歷史** — 以 SQLite 紀錄 scheduler / triggers / hotkeys / REST 的執行結果；錯誤時自動附上截圖
 - **報告產生** — 將測試紀錄匯出為 HTML、JSON 或 XML 報告，包含成功/失敗狀態
-- **MCP 伺服器** — JSON-RPC 2.0 Model Context Protocol 服務（stdio + HTTP/SSE），讓 Claude Desktop / Claude Code / 自訂 tool-use 迴圈直接驅動 AutoControl。約 90 個工具,完整協定支援(resources、prompts、sampling、roots、logging、progress、cancellation、elicitation),Bearer token 驗證 + TLS、稽核 log、rate limit、plugin 熱重載、CI fake backend
+- **MCP 伺服器** — JSON-RPC 2.0 Model Context Protocol 服務（stdio + HTTP/SSE），讓 Claude Desktop / Claude Code / 自訂 tool-use 迴圈直接驅動 AutoControl。約 100 個工具,完整協定支援(resources、prompts、sampling、roots、logging、progress、cancellation、elicitation),Bearer token 驗證 + TLS、稽核 log、rate limit、plugin 熱重載、CI fake backend。**本次新增** `ac_remote_host_start` / `ac_remote_host_stop` / `ac_remote_host_status` / `ac_remote_viewer_connect` / `ac_remote_viewer_disconnect` / `ac_remote_viewer_status` / `ac_remote_viewer_send_input` 包裝 GUI 遠端桌面分頁所用的 process-global registry,模型可以直接啟動 host、連線 viewer、轉送 mouse／keyboard／type／hotkey 動作
 - **遠端自動化** — TCP Socket 伺服器 **加上** 強化版 REST API：bearer token 認證、per-IP 速率限制 + 失敗鎖定、SQLite 稽核 hook、Prometheus `/metrics`、完整端點清單（`/health`、`/screen_size`、`/sessions`、`/screenshot`、`/execute`、`/audit/list`、`/audit/verify`、`/inspector/recent`、`/usb/devices`、`/diagnose`、…），以及 vanilla-JS 的瀏覽器 dashboard `/dashboard`（任何能 HTTP 連到主機的手機都能監看）
 - **外掛載入器** — 將定義 `AC_*` 可呼叫物的 `.py` 檔放入目錄，執行時即可註冊成 executor 指令
 - **Shell 整合** — 在自動化流程中執行 Shell 命令，支援非同步輸出擷取

docs/source/Eng/doc/mcp_server/mcp_server_doc.rst

@@ -70,6 +70,18 @@ Scheduler / triggers / hotkeys
   ``ac_hotkey_bind``, ``ac_hotkey_unbind``, ``ac_hotkey_list``,
   ``ac_hotkey_daemon_start``, ``ac_hotkey_daemon_stop``.
 
+Remote desktop (TCP host + viewer registry)
+  ``ac_remote_host_start``, ``ac_remote_host_stop``,
+  ``ac_remote_host_status``, ``ac_remote_viewer_connect``,
+  ``ac_remote_viewer_disconnect``, ``ac_remote_viewer_status``,
+  ``ac_remote_viewer_send_input``. These wrap the same singleton
+  registry the GUI's Remote Desktop tab uses, so a model can spin
+  up a host (``token``, ``bind``, ``port``, ``fps``, ``quality``,
+  ``host_id``), open a viewer to another machine, query status, and
+  forward mouse / keyboard / type / hotkey actions through the
+  active viewer. Status tools are read-only and survive
+  ``--readonly`` mode; ``send_input`` is destructive by design.
+
 Every tool carries the MCP 2025-06-18 ``annotations`` block
 (``readOnlyHint``, ``destructiveHint``, ``idempotentHint``,
 ``openWorldHint``) so well-behaved clients can auto-approve

docs/source/Eng/doc/new_features/new_features_doc.rst

@@ -673,3 +673,120 @@ upload flow.
    Keep ``trusted token holders == trusted users`` in mind, or wrap
    the headless API in your own restricted ``FileReceiver`` subclass
    that vets the destination path.
+
+
+Remote desktop — AnyDesk-style popout window
+============================================
+
+The viewer panel no longer renders the live remote screen inline —
+when the viewer authenticates, a dedicated top-level
+:class:`RemoteScreenWindow` opens with the remote desktop, and the
+panel shrinks back to the connection card + controls. Closing the
+popup ✕ disconnects the session, matching AnyDesk's session-window
+ergonomics.
+
+* New module: ``je_auto_control/gui/remote_desktop/remote_screen_window.py``
+* Wraps a ``_FrameDisplay`` and re-emits its mouse / keyboard /
+  drag-and-drop / annotation signals so the panel keeps a single
+  signal source after the popout.
+* Bottom footer carries the optional file-transfer progress label /
+  bar; hidden when no transfer is active.
+* Both the TCP ``_ViewerPanel`` and the WebRTC
+  ``_WebRTCViewerPanel`` open the popup on connect / on auth_ok and
+  close it on disconnect / on stop.
+
+Why
+   The previous layout fought for vertical space: a frame display +
+   connection card + collapsibles + action row + stats + sparklines
+   + transfer progress + status bar all stacked on one tab. Pulling
+   the live screen out into its own window leaves the operator with
+   a real workspace and keeps the control surface uncluttered.
+
+
+Remote desktop — responsive sub-tab sizing
+==========================================
+
+Every Remote Desktop sub-tab is now wrapped in a ``QScrollArea``
+with ``setWidgetResizable(True)``. The wrapper lives in
+``gui/remote_desktop/tab.py`` (helper ``_wrap_in_scroll_area``).
+
+* Small / shrunk window: a vertical scrollbar appears instead of
+  clipping the dense WebRTC panels.
+* Enlarged / 4K window: the inner panel widget grows horizontally
+  with the viewport, so the connection card and session table
+  stretch edge-to-edge instead of clustering at the top-left.
+* The bottom ``addStretch(1)`` in each panel still pushes content
+  up when there is leftover height, so the layout doesn't sag.
+
+Heavy / rarely used groups (Manual SDP, Remote Files, Sync) on the
+WebRTC viewer tab are also wrapped in collapsed-by-default
+``_CollapsibleSection`` shells via the new ``_wrap_collapsed``
+helper, halving the panel's first-paint height.
+
+Removed the previous hard ``setMaximumHeight(140)`` on the WebRTC
+host's session table: ``setMinimumHeight(140)`` keeps 140 px as a
+starting hint without capping the table on large displays.
+
+
+Remote desktop — MCP tool surface
+=================================
+
+The MCP server now wraps the same singleton remote-desktop
+registry the GUI uses. The tools live under a new
+``remote_desktop_tools()`` factory in
+``je_auto_control/utils/mcp_server/tools/_factories.py``:
+
+``ac_remote_host_start``
+   Start (or restart) the singleton TCP host with ``token``,
+   ``bind``, ``port``, ``fps``, ``quality``, ``max_clients``,
+   ``host_id``. Returns
+   ``{running, port, host_id, connected_clients}``.
+
+``ac_remote_host_stop``
+   Stop the host (no-op when nothing is running).
+
+``ac_remote_host_status``
+   Read-only snapshot of the host registry. Survives
+   ``--readonly`` mode.
+
+``ac_remote_viewer_connect``
+   Open the singleton viewer to a remote host, supporting
+   ``expected_host_id`` to verify the 9-digit ID before accepting
+   the session.
+
+``ac_remote_viewer_disconnect`` / ``ac_remote_viewer_status``
+   Close / observe the active viewer (status is read-only).
+
+``ac_remote_viewer_send_input``
+   Forward an input action dict (``mouse_move``, ``mouse_press``,
+   ``mouse_release``, ``mouse_scroll``, ``key_press``,
+   ``key_release``, ``type``, ``hotkey``) through the connected
+   viewer to the remote host. Destructive — stripped under
+   ``--readonly``.
+
+A model can now drive a complete remote-control flow without
+clicking through the GUI:
+
+.. code-block:: text
+
+   ac_remote_host_start(token="tok", bind="127.0.0.1", port=0)
+     → {"running": true, "port": 51234, "host_id": "123456789",
+        "connected_clients": 0}
+
+   # … on a different machine …
+   ac_remote_viewer_connect(host="10.0.0.5", port=51234, token="tok",
+                            expected_host_id="123456789")
+     → {"connected": true, "host_id": "123456789"}
+
+   ac_remote_viewer_send_input(action={
+       "action": "mouse_move", "x": 100, "y": 200,
+   })
+   ac_remote_viewer_send_input(action={
+       "action": "type", "text": "hello",
+   })
+
+The status / observer tools (``ac_remote_host_status``,
+``ac_remote_viewer_status``) are read-only and survive the MCP
+server's ``--readonly`` filter; everything that mutates state is
+correctly tagged ``destructiveHint: true`` so MCP clients can
+prompt for user confirmation.