dsh-sidebar-browser-control 0.1.0 → 0.2.0

package/index.js

@@ -21,8 +21,19 @@
  *
  * Only the standard library is imported: a profile-installed bundle resolves
  * outside the DSH installation, so shipped packages are not on its import path.
+ *
+ * Screenshots never touch a native Electron capture API: `webview.capturePage()`
+ * crashed the renderer (docs/architecture.md), so the Host half asks the Client
+ * half for the tab's geometry and then captures the DSH window itself with
+ * PrintWindow through PowerShell, cropping to that rectangle. It works even when
+ * the window is covered by other applications.
  */
 
+import { execFile } from 'node:child_process';
+import { mkdir, readFile, writeFile } from 'node:fs/promises';
+import { tmpdir } from 'node:os';
+import { dirname, isAbsolute, join, resolve as resolvePath } from 'node:path';
+
 /** Route prefix owned by this plugin. */
 const BASE = '/sidebar-browser-control';
 /** How long one long-poll is held open before answering with no commands. */
@@ -104,6 +115,252 @@ function admitted(req) {
   return true;
 }
 
+/**
+ * PowerShell that captures the DSH window itself (PrintWindow) and crops it.
+ *
+ * PrintWindow renders the window's own pixels, so the result does not depend on
+ * the window being foreground or unoccluded — which is exactly why it is used
+ * instead of a screen copy. The script is ASCII only and handed to
+ * `powershell.exe` as -EncodedCommand: Windows PowerShell reads .ps1 files as
+ * ANSI, so a non-ASCII script file (or `-Command` string) corrupts silently.
+ * @param outPath - PNG destination.
+ * @param crop - pixel rectangle inside the window, or zero-size for the whole window.
+ * @param restore - restore a minimized window before capturing.
+ * @returns the script source.
+ */
+export function captureScript(outPath, crop, restore = false) {
+  return [
+    '$ProgressPreference = "SilentlyContinue"',
+    'Add-Type -AssemblyName System.Drawing',
+    'Add-Type @"',
+    'using System;',
+    'using System.Runtime.InteropServices;',
+    'public class DshShot {',
+    '  [DllImport("user32.dll")] public static extern bool SetProcessDPIAware();',
+    '  [DllImport("user32.dll")] public static extern bool GetWindowRect(IntPtr hWnd, out RECT rect);',
+    '  [DllImport("user32.dll")] public static extern bool IsIconic(IntPtr hWnd);',
+    '  [DllImport("user32.dll")] public static extern bool IsWindowVisible(IntPtr hWnd);',
+    '  [DllImport("user32.dll")] public static extern bool ShowWindow(IntPtr hWnd, int command);',
+    '  [DllImport("user32.dll")] public static extern bool PrintWindow(IntPtr hWnd, IntPtr hdc, uint flags);',
+    '  [StructLayout(LayoutKind.Sequential)] public struct RECT { public int Left; public int Top; public int Right; public int Bottom; }',
+    '}',
+    '"@',
+    '[void][DshShot]::SetProcessDPIAware()',
+    '# Several processes share the app name, and a hidden helper window can be the',
+    '# first one with a handle: pick the visible, titled, un-minimised, largest one.',
+    '$best = $null',
+    '$bestScore = -1',
+    '$bestRect = $null',
+    '$bestIconic = $false',
+    "$bestTitle = ''",
+    "foreach ($candidate in (Get-Process -Name 'DeepSeek Harness' -ErrorAction SilentlyContinue)) {",
+    '  $handle = $candidate.MainWindowHandle',
+    '  if ($handle -eq 0) { continue }',
+    '  if (-not [DshShot]::IsWindowVisible($handle)) { continue }',
+    '  $rect = New-Object DshShot+RECT',
+    '  [void][DshShot]::GetWindowRect($handle, [ref]$rect)',
+    '  $area = ($rect.Right - $rect.Left) * ($rect.Bottom - $rect.Top)',
+    '  if ($area -le 0) { continue }',
+    '  $titled = $candidate.MainWindowTitle.Length -gt 0',
+    '  $iconic = [DshShot]::IsIconic($handle)',
+    '  $score = $area',
+    '  if ($titled) { $score = $score + 100000000 }',
+    '  if (-not $iconic) { $score = $score + 10000000 }',
+    '  if ($score -gt $bestScore) { $bestScore = $score; $best = $candidate; $bestRect = $rect; $bestIconic = $iconic; $bestTitle = $candidate.MainWindowTitle }',
+    '}',
+    'if ($null -eq $best) { Write-Output "error=no-window"; exit 2 }',
+    '$h = $best.MainWindowHandle',
+    'if ($bestIconic) {',
+    `  if (${restore === true ? '$true' : '$false'}) {`,
+    '    # A minimized window has no rendered surface: restore it, then re-measure.',
+    '    [void][DshShot]::ShowWindow($h, 9)',
+    '    Start-Sleep -Milliseconds 600',
+    '    [void][DshShot]::GetWindowRect($h, [ref]$bestRect)',
+    '    $bestIconic = $false',
+    '  } else {',
+    '    Write-Output "error=minimized"',
+    '    exit 4',
+    '  }',
+    '}',
+    '$r = $bestRect',
+    '$w = $r.Right - $r.Left',
+    '$t = $r.Bottom - $r.Top',
+    'if ($w -le 0 -or $t -le 0) { Write-Output "error=bad-rect"; exit 5 }',
+    '$bmp = New-Object System.Drawing.Bitmap $w, $t',
+    '$g = [System.Drawing.Graphics]::FromImage($bmp)',
+    '$hdc = $g.GetHdc()',
+    '$ok = [DshShot]::PrintWindow($h, $hdc, 2)',
+    '$g.ReleaseHdc($hdc)',
+    '$g.Dispose()',
+    'if (-not $ok) { Write-Output "error=printwindow"; exit 3 }',
+    `$cx = ${Math.round(crop.x)}`,
+    `$cy = ${Math.round(crop.y)}`,
+    `$cw = ${Math.round(crop.width)}`,
+    `$ch = ${Math.round(crop.height)}`,
+    'if ($cw -gt 0 -and $ch -gt 0) {',
+    '  if ($cx -lt 0) { $cx = 0 }',
+    '  if ($cy -lt 0) { $cy = 0 }',
+    '  if ($cx + $cw -gt $w) { $cw = $w - $cx }',
+    '  if ($cy + $ch -gt $t) { $ch = $t - $cy }',
+    '  if ($cw -gt 0 -and $ch -gt 0) {',
+    '    $c = New-Object System.Drawing.Bitmap $cw, $ch',
+    '    $g2 = [System.Drawing.Graphics]::FromImage($c)',
+    '    $g2.DrawImage($bmp, (New-Object System.Drawing.Rectangle 0, 0, $cw, $ch), (New-Object System.Drawing.Rectangle $cx, $cy, $cw, $ch), [System.Drawing.GraphicsUnit]::Pixel)',
+    '    $g2.Dispose()',
+    '    $bmp.Dispose()',
+    '    $bmp = $c',
+    '  }',
+    '}',
+    `$bmp.Save('${outPath.replace(/'/gu, "''")}', [System.Drawing.Imaging.ImageFormat]::Png)`,
+    '$bmp.Dispose()',
+    'Write-Output ("ok " + $w + "x" + $t + " title=" + $bestTitle)'
+  ].join('\n');
+}
+
+/**
+ * Run one PowerShell script without touching a file or a shell.
+ *
+ * The script is handed over as -EncodedCommand (UTF-16LE base64) so no file
+ * encoding, quoting or code page can corrupt it. CRLF line endings are required:
+ * a here-string terminator is only recognised at the start of a line, and
+ * Windows PowerShell 5.1 does not accept an LF-only terminator when the script
+ * arrives this way.
+ * @param script - ASCII script source.
+ * @param timeoutMs - kill the process after this long.
+ * @returns the trimmed standard output.
+ */
+export function runPowerShell(script, timeoutMs = 20_000) {
+  const encoded = Buffer.from(script.replace(/\r?\n/gu, '\r\n'), 'utf16le').toString('base64');
+  return new Promise((resolve, reject) => {
+    execFile('powershell.exe', ['-NoProfile', '-NonInteractive', '-ExecutionPolicy', 'Bypass', '-EncodedCommand', encoded], {
+      timeout: timeoutMs,
+      windowsHide: true,
+      maxBuffer: 8 * 1024 * 1024
+    }, (error, stdout, stderr) => {
+      if (error !== null && error !== undefined) {
+        const detail = [String(stdout || '').trim(), String(stderr || '').trim()].filter((part) => part.length > 0).join(' | ').slice(0, 400);
+        const killed = error.killed === true ? '（进程超时被终止）' : '';
+        reject(new Error(`窗口截图失败${killed}：${detail.length > 0 ? detail : error.message}`));
+        return;
+      }
+      resolve(String(stdout).trim());
+    });
+  });
+}
+
+/**
+ * Map the Client's reported page geometry to pixels inside the captured window.
+ * @param geometry - `{ rect, devicePixelRatio, outer }` from the Client half.
+ * @returns `{ x, y, width, height }` in window pixels.
+ */
+export function geometryToCrop(geometry) {
+  const dpr = Number.isFinite(geometry?.devicePixelRatio) && geometry.devicePixelRatio > 0 ? geometry.devicePixelRatio : 1;
+  const rect = geometry?.rect ?? {};
+  const outer = geometry?.outer ?? {};
+  const frameX = Number.isFinite(outer.width) && Number.isFinite(outer.innerWidth) ? Math.max(0, (outer.width - outer.innerWidth) / 2) : 0;
+  const frameY = Number.isFinite(outer.height) && Number.isFinite(outer.innerHeight) ? Math.max(0, outer.height - outer.innerHeight - frameX) : 0;
+  return {
+    x: (frameX + (Number(rect.left) || 0)) * dpr,
+    y: (frameY + (Number(rect.top) || 0)) * dpr,
+    width: (Number(rect.width) || 0) * dpr,
+    height: (Number(rect.height) || 0) * dpr
+  };
+}
+
+/**
+ * Capture the DSH window (optionally cropped to the Browser pane).
+ * @param crop - pixel rectangle from {@link geometryToCrop}.
+ * @param timeoutMs - capture timeout.
+ * @param restore - restore a minimized window first (it becomes visible again).
+ * @returns PNG bytes.
+ */
+export async function captureWindow(crop, timeoutMs = 20_000, restore = false) {
+  if (process.platform !== 'win32') {
+    throw new Error('截图目前只在 Windows 上实现（PrintWindow）。其他平台请用 browser_control 的 external 在系统浏览器里查看，或用 browser_page 的 text/elements。');
+  }
+  const outPath = join(tmpdir(), `dsh-sidebar-browser-shot-${Date.now()}.png`);
+  const output = await runPowerShell(captureScript(outPath, crop, restore), timeoutMs);
+  if (output.startsWith('error=')) {
+    const reason = output.slice('error='.length);
+    const explain = {
+      'no-window': '找不到 DSH 窗口（应用可能刚启动或已退出）。',
+      minimized: 'DSH 窗口最小化了：最小化的窗口没有可渲染表面，抓不到。请还原窗口，或带 restore: true 让插件先还原它（窗口会重新出现在屏幕上）。',
+      'bad-rect': 'DSH 窗口尺寸无效，可能正处于最小化或切换状态。',
+      printwindow: '系统拒绝了窗口抓取（PrintWindow 返回失败）。'
+    };
+    throw new Error(explain[reason] ?? `窗口截图失败：${reason}`);
+  }
+  return await readFile(outPath);
+}
+
+/**
+ * Persist a picture in the workspace and attach it to the tool result.
+ * @param ctx - plugin context.
+ * @param exec - execution whose agent owns the workspace.
+ * @param bytes - image bytes.
+ * @param label - file name stem.
+ * @param mediaType - image media type.
+ * @returns `{ path, bytes, image?, note? }`.
+ */
+async function saveScreenshot(ctx, exec, bytes, label, mediaType = 'image/png') {
+  const cwd = exec?.agent?.session?.header?.cwd;
+  const root = typeof cwd === 'string' && cwd.length > 0 ? cwd : process.cwd();
+  const stamp = new Date().toISOString().replace(/[:.]/gu, '-');
+  const safe = String(label ?? 'tab').replace(/[^A-Za-z0-9_-]/gu, '');
+  const extension = mediaType === 'image/jpeg' ? 'jpg' : mediaType === 'image/webp' ? 'webp' : mediaType === 'image/gif' ? 'gif' : 'png';
+  const relative = join('browser-shots', `${stamp}-${safe.length > 0 ? safe : 'tab'}.${extension}`);
+  const absolute = isAbsolute(relative) ? relative : resolvePath(root, relative);
+  const saved = { path: relative, bytes: bytes.length };
+  try {
+    await mkdir(dirname(absolute), { recursive: true });
+    await writeFile(absolute, bytes);
+  } catch (error) {
+    return { ...saved, note: `图片已获取但写入工作区失败：${String(error?.message ?? error)}` };
+  }
+  const attachments = ctx.get('attachments');
+  if (attachments === undefined || typeof attachments.saveImage !== 'function') return saved;
+  const attachable = ['image/png', 'image/jpeg', 'image/webp', 'image/gif'];
+  if (!attachable.includes(mediaType)) {
+    return { ...saved, note: `图片已存到工作区，但 ${mediaType} 不能作为附件显示（原图请在文件里打开）。` };
+  }
+  try {
+    const image = await attachments.saveImage({ data: bytes, mediaType, name: relative });
+    return { ...saved, image };
+  } catch (error) {
+    return { ...saved, note: `图片已保存到工作区，但无法附加到结果：${String(error?.message ?? error)}` };
+  }
+}
+
+/**
+ * Turn the Client's picture reference into bytes.
+ *
+ * The Host fetches HTTP(S) sources because the page's own `fetch` is bound by
+ * CORS; `data:` and `blob:` sources arrive already decoded.
+ * @param found - value from the Client's `image` operation.
+ * @param signal - caller cancellation.
+ * @returns `{ bytes, media_type }`.
+ */
+async function loadPicture(found, signal) {
+  const fallbackType = typeof found?.media_type === 'string' ? found.media_type : 'image/png';
+  if (typeof found?.inline === 'string') {
+    const bytes = Buffer.from(found.inline, 'base64');
+    if (bytes.length === 0) throw new Error('内联图片是空的。');
+    return { bytes, media_type: fallbackType };
+  }
+  /* `image_url` is the picture's own address; `url` stays the page's address. */
+  const url = typeof found?.image_url === 'string' ? found.image_url : undefined;
+  if (url === undefined) throw new Error('没有拿到图片地址。');
+  if (!/^https?:/u.test(url)) throw new Error(`不支持的图片来源：${url.slice(0, 80)}`);
+  const response = await fetch(url, { signal, redirect: 'follow' });
+  if (!response.ok) throw new Error(`下载图片失败：HTTP ${response.status}（该地址可能需要登录，或禁止直接访问）。`);
+  const bytes = Buffer.from(await response.arrayBuffer());
+  const limit = 8 * 1024 * 1024;
+  if (bytes.length > limit) throw new Error(`图片太大（${Math.round(bytes.length / 1024)} KB），超过 8 MB 上限。`);
+  const header = response.headers.get('content-type') ?? '';
+  const type = /^image\//u.test(header) ? header.split(';')[0] : fallbackType;
+  return { bytes, media_type: type };
+}
+
 /** One connected browser client and its pending long-poll. */
 class Client {
   /** @param id - opaque per-page identity. */
@@ -309,7 +566,11 @@ class Pilot {
 /** Render the tab list as tool text. */
 function tabsText(value) {
   const tabs = Array.isArray(value?.tabs) ? value.tabs : [];
-  const header = `platform=${value?.platform ?? 'unknown'} page_control=${value?.page_control === true ? 'yes' : 'no'} panel_open=${value?.panel_open === true ? 'yes' : 'no'}`;
+  const capture = value?.capture;
+  const captureLine = capture === undefined || capture === null
+    ? ' capture=n/a'
+    : `\ncapture: guest=${capture.guest === true ? 'yes' : 'no'} desktopCapturer=${capture.desktop_capturer === true ? 'yes' : 'no'} displayMedia=${capture.display_media === true ? 'yes' : 'no'}${Array.isArray(capture.carrier_keys) && capture.carrier_keys.length > 0 ? ` carrier=[${capture.carrier_keys.join(',')}]` : ''}`;
+  const header = `platform=${value?.platform ?? 'unknown'} page_control=${value?.page_control === true ? 'yes' : 'no'} panel_open=${value?.panel_open === true ? 'yes' : 'no'}${captureLine}`;
   if (tabs.length === 0) return `${header}\n内置浏览器当前没有打开任何标签页。`;
   const lines = tabs.map((tab) => {
     const marks = [tab.active === true ? 'active' : '', tab.on_screen === true ? 'on-screen' : 'other-session'].filter((mark) => mark.length > 0).join(',');
@@ -325,18 +586,34 @@ function actionText(value) {
   const parts = [];
   if (typeof value.action === 'string') parts.push(value.action);
   if (typeof value.tab_id === 'string') parts.push(`tab=${value.tab_id}`);
+  if (typeof value.opened_tab === 'string') parts.push(`opened=${value.opened_tab}`);
   if (typeof value.url === 'string' && value.url.length > 0) parts.push(value.url);
   if (typeof value.title === 'string' && value.title.length > 0) parts.push(`title=${value.title}`);
   const head = parts.length > 0 ? parts.join(' ') : '完成';
   const body = [];
   if (typeof value.text === 'string' && value.text.length > 0) body.push(value.text);
+  if (typeof value.opened_url === 'string' && value.opened_url.length > 0) body.push(`新标签页地址：${value.opened_url}`);
+  if (typeof value.path === 'string' && value.path.length > 0) body.push(`图片已保存：${value.path}${typeof value.width === 'number' && typeof value.height === 'number' ? `（${value.width}×${value.height}）` : ''}`);
   if (typeof value.note === 'string' && value.note.length > 0) body.push(`注意：${value.note}`);
   if (value.value !== undefined && value.value !== null) body.push(JSON.stringify(value.value));
   if (Array.isArray(value.links) && value.links.length > 0) {
     body.push(value.links.map((link) => `- ${link.text ?? ''} ${link.href ?? ''}`).join('\n'));
   }
+  if (Array.isArray(value.images) && value.images.length > 0) {
+    body.push(value.images.map((picture) => `- [${picture.index ?? '?'}] ${picture.tag ?? 'img'} ${picture.width ?? '?'}×${picture.height ?? '?'}${picture.alt !== undefined ? ` alt="${picture.alt}"` : ''} ${picture.src ?? ''}`).join('\n'));
+  }
   if (Array.isArray(value.elements) && value.elements.length > 0) {
-    body.push(value.elements.map((element) => `- ${element.selector ?? ''} ${element.text ?? ''}`).join('\n'));
+    body.push(value.elements.map((element) => {
+      const meta = [
+        element.type !== undefined && element.type !== element.tag ? `type=${element.type}` : '',
+        element.name !== undefined ? `name=${element.name}` : '',
+        element.autocomplete !== undefined ? `autocomplete=${element.autocomplete}` : '',
+        element.checked === true ? 'checked' : ''
+      ].filter((part) => part.length > 0).join(' ');
+      const label = element.label !== undefined && element.label.length > 0 ? ` “${element.label}”` : '';
+      const href = typeof element.href === 'string' && element.href.length > 0 ? ` (${element.href})` : '';
+      return `- ${element.selector ?? ''}${meta.length > 0 ? ` [${meta}]` : ''} — ${element.text ?? ''}${label}${href}`;
+    }).join('\n'));
   }
   return body.length === 0 ? head : `${head}\n${body.join('\n')}`;
 }
@@ -464,21 +741,45 @@ export function apply(ctx) {
 
   ctx.effect(() => ctx.tools.register({
     name: 'browser_page',
-    description: "Read and act on the page inside DSH's built-in browser tab (Desktop app only; browser_tabs reports page_control). text returns the page's visible text, links lists anchors, elements lists interactive elements with a usable selector for click, eval runs JavaScript in the page and returns its JSON value, click clicks a selector or the element containing text, type fills a field (optionally submitting), scroll scrolls, wait waits for a selector or text.",
+    description: "Read and act on the page inside DSH's built-in browser tab (Desktop app only; browser_tabs reports page_control). text returns the page's visible text, links lists anchors, elements lists interactive elements (with type/name/label/placeholder/autocomplete for form fields) and a usable selector, eval runs JavaScript and returns its JSON value, click clicks a selector or the element containing text, type fills one field, fill fills several fields by label/name/placeholder/selector and can submit, key presses Enter/Tab/Escape, login finds the account and password fields (including inside a same-origin iframe), fills and submits them, scroll scrolls, wait waits for a selector, text or url_contains, image returns a picture from the page so you can look at it, screenshot captures the browser pane itself with vision. A link that opens a new window becomes a NEW sidebar tab: click reports it as opened_tab, and you must use that id afterwards. A page that failed to load is reported as such instead of returning empty content. Passwords are never echoed in a result.",
     parameters: {
       type: 'object',
       additionalProperties: false,
       properties: {
         action: {
           type: 'string',
-          enum: ['text', 'links', 'elements', 'eval', 'click', 'type', 'scroll', 'wait'],
+          enum: ['text', 'links', 'elements', 'eval', 'click', 'type', 'fill', 'key', 'login', 'scroll', 'wait', 'image', 'images', 'screenshot'],
           description: 'What to do with the page.'
         },
         tab_id: { type: 'string', description: 'Target tab id from browser_tabs; omit for the active tab.' },
-        selector: { type: 'string', description: 'CSS selector for click/type/scroll/wait.' },
+        selector: { type: 'string', description: 'CSS selector for click/type/scroll/wait/image.' },
         text: { type: 'string', description: 'Text to type, or text to match for click/wait when no selector is given.' },
         code: { type: 'string', description: 'JavaScript for eval. Use an expression, or statements ending in a return.' },
-        submit: { type: 'boolean', description: 'type: submit the form afterwards.' },
+        fields: {
+          type: 'array',
+          description: 'fill: one entry per field. Each entry matches a field by selector, or by label/name/placeholder/type, and carries the value to set.',
+          items: {
+            type: 'object',
+            additionalProperties: false,
+            properties: {
+              selector: { type: 'string' },
+              label: { type: 'string' },
+              name: { type: 'string' },
+              placeholder: { type: 'string' },
+              type: { type: 'string' },
+              value: { type: 'string' }
+            }
+          }
+        },
+        user: { type: 'string', description: 'login: account name, email or phone to fill in.' },
+        password: { type: 'string', description: 'login: password to fill in. It is used once and never returned.' },
+        user_selector: { type: 'string', description: 'login: explicit CSS selector for the account field.' },
+        password_selector: { type: 'string', description: 'login: explicit CSS selector for the password field.' },
+        key: { type: 'string', description: 'key: Enter (default), Tab, Escape, ArrowDown…' },
+        index: { type: 'number', description: 'image: 1-based index from the images action.' },
+        restore: { type: 'boolean', description: 'screenshot: restore the window first when it is minimized (it becomes visible again). Refused by default, because the capture cannot render a minimized window.' },
+        submit: { type: 'boolean', description: 'type/fill: submit the form afterwards.' },
+        url_contains: { type: 'string', description: 'wait: wait until the address contains this text.' },
         ms: { type: 'number', description: 'wait: timeout in milliseconds (default 10000). scroll: pixels to scroll (default one viewport).' },
         max_chars: { type: 'number', description: 'text/links: truncate the returned text (default 8000 characters).' }
       },
@@ -493,12 +794,47 @@ export function apply(ctx) {
           url: { type: 'string' },
           title: { type: 'string' },
           text: { type: 'string' },
-          note: { type: 'string' }
+          note: { type: 'string' },
+          path: { type: 'string' },
+          bytes: { type: 'number' },
+          image: { type: 'object' }
         }
       },
-      render: (_args, value) => [{ type: 'text', text: actionText(value) }]
+      render: (_args, value) => {
+        const blocks = [{ type: 'text', text: actionText(value) }];
+        if (value?.image !== undefined && value.image !== null) blocks.push({ type: 'image', attachment: value.image });
+        return blocks;
+      }
     },
     async execute(args, exec) {
+      if (args?.action === 'screenshot') {
+        const geometry = await pilot.call('geometry', { tab_id: args.tab_id }, exec.signal);
+        const crop = geometryToCrop(geometry);
+        const restore = args?.restore === true;
+        const bytes = await captureWindow(crop, 20_000, restore);
+        const saved = await saveScreenshot(ctx, exec, bytes, geometry?.tab_id ?? 'browser');
+        return {
+          action: 'screenshot',
+          tab_id: geometry?.tab_id,
+          url: geometry?.url,
+          title: geometry?.title,
+          width: Math.round(crop.width),
+          height: Math.round(crop.height),
+          restored: restore ? true : undefined,
+          ...saved,
+          note: saved.note ?? (restore
+            ? '这是内置浏览器面板本身的截图，已附在结果里；截图前窗口是被还原过的。'
+            : '这是内置浏览器面板本身的截图，已附在结果里。')
+        };
+      }
+      if (args?.action === 'image') {
+        const found = await pilot.call('image', args, exec.signal);
+        const loaded = await loadPicture(found, exec.signal);
+        const named = { ...found, ...loaded };
+        const saved = await saveScreenshot(ctx, exec, loaded.bytes, `image-${found.width ?? 0}x${found.height ?? 0}`);
+        const { bytes: _bytes, ...rest } = named;
+        return { ...rest, ...saved, note: saved.note ?? (found.alt !== undefined ? `页面里的图片：${found.alt}` : '页面里的图片，已附在结果里。') };
+      }
       return await pilot.call('page', args, exec.signal);
     }
   }), 'sidebar-browser-control.tool.page');
@@ -506,8 +842,8 @@ export function apply(ctx) {
   ctx.inject(['commands'], (scope) => {
     scope.effect(() => scope.commands.register({
       name: 'browser',
-      description: '驱动 DSH 内置浏览器：/browser open <url> | tabs | back | forward | reload | close | expand | collapse | fullscreen | split | read',
-      input: { hint: 'open <url> | tabs | back | forward | reload | close | expand | collapse | fullscreen | split | read' },
+      description: '驱动 DSH 内置浏览器：/browser open <url> | tabs | back | forward | reload | close | expand | collapse | fullscreen | split | read | images | image <N> | screenshot',
+      input: { hint: 'open <url> | tabs | back | forward | reload | close | expand | collapse | fullscreen | split | read | images | image <N> | screenshot' },
       async handler(invocation) {
         try {
           return { kind: 'success', text: await runCommand(pilot, invocation) };
@@ -541,6 +877,21 @@ async function runCommand(pilot, invocation) {
     }
     case 'read':
       return actionText(await pilot.call('page', { action: 'text', ...argument.length > 0 ? { tab_id: argument } : {} }, signal));
+    case 'screenshot': {
+      const geometry = await pilot.call('geometry', argument.length > 0 ? { tab_id: argument } : {}, signal);
+      const crop = geometryToCrop(geometry);
+      const bytes = await captureWindow(crop);
+      const saved = await saveScreenshot(ctx, { agent: invocation.agent }, bytes, geometry?.tab_id ?? 'browser');
+      return actionText({ action: 'screenshot', tab_id: geometry?.tab_id, url: geometry?.url, title: geometry?.title, width: Math.round(crop.width), height: Math.round(crop.height), ...saved });
+    }
+    case 'image': {
+      const found = await pilot.call('image', { index: Number.parseInt(argument, 10) || 1 }, signal);
+      const loaded = await loadPicture(found, signal);
+      const saved = await saveScreenshot(ctx, { agent: invocation.agent }, loaded.bytes, `image-${found.width ?? 0}x${found.height ?? 0}`, loaded.media_type);
+      return actionText({ ...found, ...saved });
+    }
+    case 'images':
+      return actionText(await pilot.call('page', { action: 'images', ...argument.length > 0 ? { tab_id: argument } : {} }, signal));
     case 'back':
     case 'forward':
     case 'reload':

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "dsh-sidebar-browser-control",
-  "version": "0.1.0",
-  "description": "Drive DSH's own built-in browser (the right-hand Sidebar's Browser tab) from the agent: list tabs, open, navigate, back/forward, reload, read the page, run JavaScript, click and type. No second Chromium process.",
+  "version": "0.2.0",
+  "description": "Drive DSH's own built-in browser (the right-hand Sidebar's Browser tab) from the agent: list tabs, open, navigate, read the page, run JavaScript, click, fill forms, log in, and look at pictures or the pane itself with vision. No second Chromium process.",
   "keywords": [
     "dsh",
     "dsh-plugin",