@hope_phenom/dsh-plugin-remote-use 0.0.0-stage → 0.2.24

package/lib/tools/selfcheck-fix-prompt.md

@@ -0,0 +1,206 @@
+# 从机环境自动配置提示词（复制给从机的 DSH 会话）
+
+> 用法：在**从机**的 DSH 会话里新建一个对话，把本文件从「你要做什么」往下整段粘进去。
+> 它让从机自己的 agent 去跑自检、按项修复、再复跑证明。需要管理员权限或需要人登录/解锁的步骤，
+> 它会停下来告诉你，而不是假装成功。
+
+---
+
+## 你要做什么
+
+这台 Windows 机器是 DSH 的**远程工作节点**。请实际执行下面的清单，把它的 GUI 就绪环境配好，并用自检脚本
+**证明**每一项都过了。不要猜测；每一步都贴**原始输出**；做不到的项明确写"做不到，原因/需要什么"。
+
+## 纪律
+
+- 只在这台机器上操作；**不要修改 DSH 安装目录，也不要手改任何 profile**。
+- 需要管理员权限的步骤先说明再执行；安装驱动/软件前先征得操作者同意。
+- 装驱动只用官方发布（Virtual-Display-Driver 是 MIT）；**不要**为了"让检查过"而关 UAC / 防火墙 / 安全策略。
+- 最后必须**复跑自检并贴 JSON**。"命令返回 0" 不等于 "检查项通过"。
+
+## 第 0 步：找到并运行自检
+
+```powershell
+$selfcheck = Get-ChildItem "$env:USERPROFILE\.dsh\profiles\*\node_modules\@hope_phenom\dsh-plugin-remote-use\slave\selfcheck.ps1" -ErrorAction SilentlyContinue | Select-Object -First 1
+if ($null -eq $selfcheck) { throw "找不到 selfcheck.ps1：这台机器还没装 @hope_phenom/dsh-plugin-remote-use" }
+$ps = if (Get-Command pwsh -ErrorAction SilentlyContinue) { "pwsh" } else { "powershell" }
+& $ps -NoProfile -ExecutionPolicy Bypass -File $selfcheck.FullName
+Get-Content (Join-Path $selfcheck.DirectoryName "selfcheck.json") -Raw
+```
+
+- 自检必须在**已登录、已解锁的交互式桌面会话**里跑（就是 DSH 桌面进程所在的那个会话）。
+  SSH / 计划任务 / 服务里跑会得到 `session=0` 的假阴性（S1 显示 0 块屏、S6/S7/S8 n/a），那些不是真问题。
+- 布防时填了任务目录的话，加 `-Workspace "<任务目录>"`（它会顺带测量 S20 工作目录完整性）。
+- 把输出的 `verdict`、`blockingFailures` 和 `checks[]`（每项的 `id/status/actual/fix`）都留着，后面要对比。
+
+## 第 1 步：按报告逐项修复
+
+> 先修 **blocking** 项，再修 warning。每修完一类就复跑一次自检，确认那一项真的变了。
+
+### S1 / S2 显示适配器与虚拟显示器（blocking）
+
+判据：至少一个 `ATTACHED_TO_DESKTOP` 的适配器；并且存在**带 EDID 的虚拟显示器**，不能回落到 `WinDisc`
+（Microsoft 基本显示适配器，通常 1024x768）。
+
+```powershell
+Get-PnpDevice -Class Display | Select-Object Status, Problem, FriendlyName, InstanceId | Format-List
+Add-Type -AssemblyName System.Windows.Forms
+[System.Windows.Forms.Screen]::AllScreens | ForEach-Object { "$($_.DeviceName) $($_.Bounds.Width)x$($_.Bounds.Height) at $($_.Bounds.X),$($_.Bounds.Y) primary=$($_.Primary)" }
+```
+
+- 如果 VDD 已装但被禁用（`Problem=CM_PROB_DISABLED` / Code 22）：
+  `Enable-PnpDevice -InstanceId 'ROOT\DISPLAY\0000' -Confirm:$false`（需要管理员）。
+- 如果根本没装 VDD：按官方 **Virtual-Display-Driver（VDD，MIT）** 的发布说明安装（需要管理员，通常要重启）。
+  **先向操作者确认**再装；装完再跑一次自检，S2 的 `actual` 里应出现 VDD 与 EDID，而不是 `WinDisc`。
+- ⚠️ **不要把虚拟屏设成主屏**：Windows 的登录界面只画在主屏上，主屏若是没有物理输出的虚拟屏，
+  操作者会看到黑屏、无法登录。主屏保持物理屏，虚拟屏作为**第二块**。
+
+### S3 / S4 分辨率与原点（S3 警告，S4 阻断）
+
+- **S4 是阻断项**：主屏必须在 `(0,0)`，否则 GUI 自动化的绝对坐标会整体偏移（而且不报错）。
+- S3 期望主屏 `1920x1080`；物理主屏做不到时它只是警告，不用强求。
+
+下面的脚本会把**主屏**设成 1920x1080@(0,0)，并（可选）把第二块屏设成 1920x1080@(1920,0)：
+
+```powershell
+Add-Type @"
+using System;
+using System.Runtime.InteropServices;
+public class DispCfg {
+  [StructLayout(LayoutKind.Sequential, CharSet = CharSet.Ansi)]
+  public struct DEVMODE {
+    [MarshalAs(UnmanagedType.ByValTStr, SizeConst = 32)] public string dmDeviceName;
+    public short dmSpecVersion; public short dmDriverVersion; public short dmSize; public short dmDriverExtra;
+    public int dmFields;
+    public int dmPositionX; public int dmPositionY; public int dmDisplayOrientation; public int dmDisplayFixedOutput;
+    public short dmColor; public short dmDuplex; public short dmYResolution; public short dmTTOption; public short dmCollate;
+    [MarshalAs(UnmanagedType.ByValTStr, SizeConst = 32)] public string dmFormName;
+    public short dmLogPixels;
+    public int dmBitsPerPel; public int dmPelsWidth; public int dmPelsHeight;
+    public int dmDisplayFlags; public int dmDisplayFrequency;
+    public int dmICMMethod; public int dmICMIntent; public int dmMediaType; public int dmDitherType;
+    public int dmReserved1; public int dmReserved2; public int dmPanningWidth; public int dmPanningHeight;
+  }
+  [DllImport("user32.dll", CharSet = CharSet.Ansi)] public static extern bool EnumDisplaySettings(string dev, int mode, ref DEVMODE dm);
+  [DllImport("user32.dll", CharSet = CharSet.Ansi)] public static extern int ChangeDisplaySettingsEx(string dev, ref DEVMODE dm, IntPtr hwnd, int flags, IntPtr param);
+  public const int ENUM_CURRENT_SETTINGS = -1;
+  public const int CDS_UPDATEREGISTRY = 0x01;
+  public const int DM_POSITION = 0x20, DM_PELSWIDTH = 0x80000, DM_PELSHEIGHT = 0x100000, DM_DISPLAYFREQUENCY = 0x400000;
+}
+"@
+
+Add-Type -AssemblyName System.Windows.Forms
+$displays = @([System.Windows.Forms.Screen]::AllScreens)
+
+function Set-DisplayMode([string]$dev, [int]$w, [int]$h, [int]$x, [int]$y, [int]$hz) {
+  $dm = New-Object DispCfg+DEVMODE
+  $dm.dmSize = [short][Runtime.InteropServices.Marshal]::SizeOf($dm)
+  if (-not [DispCfg]::EnumDisplaySettings($dev, -1, [ref]$dm)) { return "READ-FAILED $dev" }
+  $dm.dmPelsWidth = $w; $dm.dmPelsHeight = $h
+  $dm.dmPositionX = $x; $dm.dmPositionY = $y
+  $dm.dmDisplayFrequency = $hz
+  $dm.dmFields = [DispCfg]::DM_PELSWIDTH -bor [DispCfg]::DM_PELSHEIGHT -bor [DispCfg]::DM_POSITION -bor [DispCfg]::DM_DISPLAYFREQUENCY
+  $rc = [DispCfg]::ChangeDisplaySettingsEx($dev, [ref]$dm, [IntPtr]::Zero, [DispCfg]::CDS_UPDATEREGISTRY, [IntPtr]::Zero)
+  return "rc=$rc $dev ${w}x${h}@${x},${y} ${hz}Hz"
+}
+
+$primary = $displays | Where-Object { $_.Primary } | Select-Object -First 1
+Set-DisplayMode $primary.DeviceName 1920 1080 0 0 60
+# 可选：把第二块屏排到主屏右侧，同样 1920x1080
+$others = @($displays | Where-Object { -not $_.Primary })
+if ($others.Count -gt 0) { Set-DisplayMode $others[0].DeviceName 1920 1080 1920 0 60 }
+Start-Sleep -Seconds 2
+[System.Windows.Forms.Screen]::AllScreens | ForEach-Object { "$($_.DeviceName) $($_.Bounds.Width)x$($_.Bounds.Height) at $($_.Bounds.X),$($_.Bounds.Y) primary=$($_.Primary)" }
+```
+
+返回码含义：`0` 成功、`1` 需要重启、`-1` 失败、`-2` 不支持该模式。失败就贴返回码，**不要换别的分辨率硬试**。
+（如果该屏不支持 1920x1080，S3 保持警告即可；S4 必须满足。）
+
+### S6 / S7 / S8 交互式会话（S6/S7/S8）
+
+判据：自检运行在 **console 会话**里，且该会话**已登录、未锁屏**；S7/S8 需要真的能动鼠标（S8 会移动光标）。
+
+- 这三项**不是 agent 远程能修的**：锁屏/未登录时必须由操作者在这台机器上（或 RDP）登录并解锁。
+- 先用 `query session` 确认有一个 `Active` 的 console 会话；S6 会告诉你 `mySession` 与 `consoleId` 是否一致。
+- 如果自检是在 SSH 里跑的，换到 DSH 桌面会话里重跑（见第 0 步）。
+
+### S9 睡眠/休眠超时（blocking）
+
+```powershell
+powercfg /change standby-timeout-ac 0
+powercfg /change standby-timeout-dc 0
+powercfg /change hibernate-timeout-ac 0
+powercfg /change hibernate-timeout-dc 0
+powercfg /query SCHEME_CURRENT SUB_SLEEP STANDBYIDLE
+```
+
+可能需要管理员。目标是 AC/DC 的超时都变成 `0`。
+
+### S10 屏保（warning）
+
+```powershell
+$k = "HKCU:\Control Panel\Desktop"
+Set-ItemProperty $k -Name ScreenSaveActive  -Value "0" -Type String
+Set-ItemProperty $k -Name ScreenSaveTimeOut -Value "0" -Type String
+Set-ItemProperty $k -Name ScreenSaverIsSecure -Value "0" -Type String
+```
+
+### S11 自动锁屏策略（blocking）
+
+```powershell
+$k = "HKLM:\SOFTWARE\Microsoft\Windows\CurrentVersion\Policies\System"
+if (Test-Path $k) { Remove-ItemProperty $k -Name InactivityTimeoutSecs -ErrorAction SilentlyContinue }
+```
+
+需要管理员；删完注销/重启才彻底生效。
+
+### S12 控制台 QuickEdit（blocking）
+
+`QuickEdit` 会让控制台在选中文字时**挂起输出**，GUI 任务里表现为"命令卡住"。对当前用户和默认控制台都关掉：
+
+```powershell
+Set-ItemProperty "HKCU:\Console" -Name QuickEdit -Value 0 -Type DWord
+Get-ChildItem "HKCU:\Console" -ErrorAction SilentlyContinue | ForEach-Object {
+  Set-ItemProperty $_.PSPath -Name QuickEdit -Value 0 -Type DWord -ErrorAction SilentlyContinue
+}
+```
+
+**新开的控制台**才会读到新值（中继/自检下次启动时生效）。
+
+### S14 / S15 DSH 与模型凭据（blocking）
+
+- S14：DSH 已安装（`actual` 会给出安装版本）。
+- S15：`%USERPROFILE%\.dsh\.credentials.yaml` 存在。**缺凭据只能由人登录 DSH 解决**（agent 不能替登录）。
+
+### S16 磁盘空间（blocking）
+
+自检要求 ≥ 50 GB 可用。清理临时目录/旧截图（`%TEMP%\dsh-plugin-remote-use\shots` 等），或在布防面板把
+截图缓存上限设小。
+
+### S18 被测应用（warning，可选）
+
+只有在需要从机验证某个具体程序时才配：给自检传 `-App "<exe 路径>"`，或设置 `DSH_SELFCHECK_APP`。
+没配就是"不验证具体应用"，不影响 GUI 任务。
+
+### S20 工作目录完整性（warning）
+
+`danger-full-access` 下中继启动时会自动修复任务目录的 Low 完整性标签。需要手工修时：
+
+```powershell
+powershell -NoProfile -ExecutionPolicy Bypass -File (Join-Path $selfcheck.DirectoryName "repair-workspace-integrity.ps1") -Path "<任务目录>"
+```
+
+## 第 2 步：复跑并汇报
+
+再跑一次第 0 步（带 `-Workspace`），然后把**修复前/修复后**两份 `verdict + checks` 贴出来，逐项说明：
+
+- 哪些从 fail → pass；
+- 哪些仍是 fail，原因是什么、需要操作者做什么（登录/解锁、管理员、装驱动、重启）；
+- **不要**把"我改了注册表"当成"检查通过"——以复跑结果为准。
+
+## 不要做的事
+
+- 不要把虚拟屏设成主屏（登录界面会黑屏）。
+- 不要关 UAC / 防火墙 / Defender 来"让检查通过"。
+- 不要在 SSH/服务会话里下结论（你会看到 session=0 的假阴性）。
+- 不要改 DSH 的安装目录或 profile 文件；本清单只动操作系统设置。

package/locale/en.json

@@ -0,0 +1,6 @@
+{
+  "meta": {
+    "title": "Remote Worker (LAN)",
+    "description": "Delegate tasks to a paired LAN worker and drive it like a subagent over ACP."
+  }
+}

package/locale/zh.json

@@ -0,0 +1,6 @@
+{
+  "meta": {
+    "title": "远程工作节点（局域网）",
+    "description": "把任务委托给已配对的局域网从机，并用 ACP 像调用子 Agent 一样驱动它。"
+  }
+}

package/package.json

@@ -1,6 +1,74 @@
 {
   "name": "@hope_phenom/dsh-plugin-remote-use",
-  "version": "0.0.0-stage",
-  "stub": true,
-  "description": "Temporary package placeholder for staged publishing"
-}
+  "version": "0.2.24",
+  "description": "Delegate DSH tasks to a paired LAN worker and drive it like a subagent (ACP over WebSocket).",
+  "license": "MIT",
+  "author": "Hope.Phenom <hope.phenom@foxmail.com>",
+  "homepage": "https://github.com/Hope-Phenom/dsh-plugin-remote-use#readme",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/Hope-Phenom/dsh-plugin-remote-use.git",
+    "directory": "packages/dsh-plugin-remote-use"
+  },
+  "bugs": {
+    "url": "https://github.com/Hope-Phenom/dsh-plugin-remote-use/issues"
+  },
+  "keywords": [
+    "dsh",
+    "dsh-plugin",
+    "acp",
+    "agent-client-protocol",
+    "subagent",
+    "remote",
+    "lan",
+    "gui-automation"
+  ],
+  "type": "module",
+  "main": "lib/index.js",
+  "icon": "./assets/icon.svg",
+  "exports": {
+    ".": "./lib/index.js",
+    "./client": "./lib/client.js",
+    "./package.json": "./package.json",
+    "./locale/*.json": "./locale/*.json"
+  },
+  "files": [
+    "lib",
+    "assets",
+    "locale",
+    "cordis.patch.yml",
+    "README.md",
+    "README.zh.md",
+    "slave/*.mjs",
+    "slave/*.ps1",
+    "slave/acp-profile-approval.patch.yml",
+    "slave/node_modules"
+  ],
+  "scripts": {
+    "test": "node --test"
+  },
+  "dsh": {
+    "bundle": {
+      "patch": "./cordis.patch.yml"
+    },
+    "client": {
+      "platform": "web"
+    }
+  },
+  "dependencies": {
+    "@agentclientprotocol/sdk": "1.4.0",
+    "ws": "8.22.0",
+    "zod": "4.6.5"
+  },
+  "peerDependencies": {
+    "@deepseek-ai/schemastery": "*"
+  },
+  "peerDependenciesMeta": {
+    "@deepseek-ai/schemastery": {
+      "optional": true
+    }
+  },
+  "engines": {
+    "node": ">=20"
+  }
+}

package/slave/acp-profile-approval.patch.yml

@@ -0,0 +1,66 @@
+# Worker-side ACP profile patch: the approval row, made safe for a full-access deployment.
+#
+# WHERE THIS GOES: `%DSH_HOME%\profiles\acp\cordis.patch.yml` on the worker, as the entry for the
+# `approval` row. The rest of that file is machine-specific (model pin, sandbox mediator paths) and
+# is not reproduced here -- only the one entry that decides whether full access is SAFE to set.
+#
+# WHEN YOU NEED IT: when this worker's launch spec is generated with
+#   node make-launch-spec.mjs --permission-mode danger-full-access
+# i.e. when the machine is a DEDICATED GUI worker. That is the intended setup for GUI automation
+# (see DESIGN section 8.6.3 and PITFALLS F33): a per-call escalation cannot host a multi-step GUI
+# session, so the permission level is a property of the deployment, not of each task.
+#
+# ---------------------------------------------------------------------------
+# WHY THIS EXISTS (PITFALLS F34 -- read it before changing this)
+#
+# `dsh-base` derives BOTH the sandbox mode and the approval policy from ONE environment variable:
+#
+#   cordis.patch.yml:232   sandbox-policy.mode  = process.env.DSH_PERMISSION_MODE ?? 'workspace-write'
+#   cordis.patch.yml:248   approval.policy      = (that value === 'danger-full-access') ? 'never' : 'ask'
+#
+# So DSH pairs "full-access sandbox" with "approval: never" -- that pairing IS the `danger-full-access`
+# PRESET. The `permission` row (`@deepseek-ai/dsh-permission-presets`) is composed with a fixed table:
+#
+#   read-only           -> sandbox: read-only,          approval: ask
+#   workspace-write     -> sandbox: workspace-write,    approval: ask
+#   danger-full-access  -> sandbox: danger-full-access, approval: never
+#
+# and NO `defaultPreset`. Its schema says `defaultPreset` is "required when composition defaults match
+# no preset".
+#
+# A worker profile patch that hard-codes `policy: ask` therefore breaks the pairing: with
+# `DSH_PERMISSION_MODE=danger-full-access` it composes to (danger-full-access, ask), which matches NO
+# preset, leaves `defaultPreset` unset and unsatisfiable, and can stop the `permission` service from
+# loading -- i.e. the ACP process never comes up and the worker becomes UNREACHABLE. That is the F24
+# failure shape, reached through a configuration combination rather than through code.
+#
+# A hard-coded `ask` was the RIGHT call when it was made (F18's round: it turns a sandbox refusal into
+# a request the master can see, instead of a silent failure). It is retained here for exactly those
+# cases -- the expression below still yields `ask` for every mode except full access. What changes is
+# only that the profile stops DISAGREEING with the base composition about the full-access pairing.
+#
+# The consequence must be stated plainly rather than buried: A FULL-ACCESS WORKER ALSO RUNS WITH
+# APPROVAL POLICY `never`. That is DSH's own preset semantics and it is coherent -- with an
+# unrestricted sandbox there is no refusal left for the approval queue to surface. It is not a bypass
+# of the queue; it is the queue having nothing to ask about. Every other mode keeps `ask`, and so
+# does a deployment that sets no mode at all.
+# ---------------------------------------------------------------------------
+
+- id: approval
+  name: '@deepseek-ai/dsh-user-approval'
+  config:
+    # MUST stay this expression, not a literal. It mirrors `dsh-base` byte for byte for the modes that
+    # yield `ask`, so this entry is a no-op except when a full-access deployment is in force.
+    #
+    # THE DOUBLE QUOTES ARE LOAD-BEARING, and leaving them out is not a style choice. The expression
+    # contains ": " (before 'ask' in the ternary), and YAML reads an unquoted ": " as a MAPPING rather
+    # than as part of a scalar -- so the unquoted form parses as the object {"...? 'never'": "'ask'"},
+    # and Cordis rejects it: `$.policy expected "ask" | "never" but got {"[object Object]":"ask"}`.
+    # `dsh-base` quotes it for exactly this reason (cordis.patch.yml:248).
+    #
+    # MEASURED 2026-10-15: with the quotes missing, TWO entries failed to activate (approval failed
+    # validation; permission-presets then sat "pending (waiting for service: approval)") while the
+    # agent still answered and the sandbox mode still applied -- so nothing looked broken unless
+    # stderr was read. A guard that checks the expression's tokens does NOT catch this; only running
+    # it, or checking the quoting, does.
+    policy: !!js "(process.env.DSH_PERMISSION_MODE ?? 'workspace-write') === 'danger-full-access' ? 'never' : 'ask'"