diff --git a/.gitignore b/.gitignore
index f6d70ea..ba1e6e9 100644
--- a/.gitignore
+++ b/.gitignore
@@ -12,3 +12,5 @@ test-output/
 web-capability-audit.txt
 
 *.traineddata
+
+temp/
diff --git a/electron-builder.config.mjs b/electron-builder.config.mjs
index 24716dc..14c707d 100644
--- a/electron-builder.config.mjs
+++ b/electron-builder.config.mjs
@@ -23,6 +23,8 @@ export default {
   extraResources: [
     { from: path.join(root, "resources", "catalog.json"), to: "catalog.json" },
     { from: path.join(root, "resources", "excluded.json"), to: "excluded.json" },
+    { from: path.join(root, "resources", "desktop-picker.ps1"), to: "desktop-picker.ps1" },
+    { from: path.join(root, "resources", "models"), to: "models" },
     { from: path.join(root, "resources", "fixture"), to: "fixture" }
   ],
   asar: true,
diff --git a/electron.vite.config.ts b/electron.vite.config.ts
index 9d4947b..ef3260b 100644
--- a/electron.vite.config.ts
+++ b/electron.vite.config.ts
@@ -12,7 +12,7 @@ export default defineConfig({
     },
     build: {
       rollupOptions: {
-        input: { index: resolve("src/main/index.ts") }
+        input: { index: resolve("src/main/index.ts"), "local-vision/ui-detect-worker": resolve("src/main/local-vision/ui-detect-worker.ts") }
       }
     }
   },
diff --git a/package-lock.json b/package-lock.json
index 96e6e25..d7b6aca 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -22,6 +22,7 @@
         "mssql": "^12.7.13",
         "mysql2": "^3.24.6",
         "nodemailer": "^6.9.16",
+        "onnxruntime-node": "^1.31.1",
         "pdf-lib": "^1.17.1",
         "pdfjs-dist": "^4.10.38",
         "pg": "^8.23.1",
@@ -2362,7 +2363,7 @@
     },
     "node_modules/@napi-rs/canvas/node_modules/@napi-rs/canvas-linux-riscv64-gnu": {
       "version": "1.1.0",
-      "resolved": "https://registry.npmmirror.com/@napi-rs/canvas-linux-riscv64-gnu/-/canvas-linux-riscv64-gnu-1.1.0.tgz",
+      "resolved": "https://registry.npmjs.org/@napi-rs/canvas-linux-riscv64-gnu/-/canvas-linux-riscv64-gnu-1.1.0.tgz",
       "integrity": "sha512-kJ7m5n2VffxWXM5QGa5lfq8ZN4LdOJEoz3aAJAmIqYvKiZw1gRdXqpucnRXsF6D9nLYiiIlevaAtAclR7c5ilg==",
       "cpu": [
         "riscv64"
@@ -8452,7 +8453,7 @@
     },
     "node_modules/onnxruntime-node": {
       "version": "1.31.1",
-      "resolved": "https://registry.npmmirror.com/onnxruntime-node/-/onnxruntime-node-1.31.1.tgz",
+      "resolved": "https://registry.npmjs.org/onnxruntime-node/-/onnxruntime-node-1.31.1.tgz",
       "integrity": "sha512-ijnSghbDHVWaQ2nDPmVccPlCHI8kH1Mx1aJ4a/qzRP16bBcVs4E3EqZQKZCkhWnbqN1tuQWxagimFtDpbPYYQw==",
       "hasInstallScript": true,
       "license": "MIT",
diff --git a/package.json b/package.json
index d5d0e35..d73ed34 100644
--- a/package.json
+++ b/package.json
@@ -31,6 +31,7 @@
     "mssql": "^12.7.13",
     "mysql2": "^3.24.6",
     "nodemailer": "^6.9.16",
+    "onnxruntime-node": "^1.31.1",
     "pdf-lib": "^1.17.1",
     "pdfjs-dist": "^4.10.38",
     "pg": "^8.23.1",
diff --git a/resources/desktop-picker.ps1 b/resources/desktop-picker.ps1
index 43c5def..4f45246 100644
--- a/resources/desktop-picker.ps1
+++ b/resources/desktop-picker.ps1
@@ -150,10 +150,15 @@ function Set-ClickThrough($form) {
 }
 
 function Set-HighlightBox([int]$x, [int]$y, [int]$w, [int]$h) {
-  $hl.Location = New-Object System.Drawing.Point(($x - 2), ($y - 2))
-  $hl.Size = New-Object System.Drawing.Size(($w + 4), ($h + 4))
-  $hl.Show()
-  Set-ClickThrough $hl
+  $bounds = New-Object System.Drawing.Rectangle(($x - 2), ($y - 2), ($w + 4), ($h + 4))
+  if (-not $hl.Visible) {
+    $hl.Bounds = $bounds
+    $hl.Show()
+    Set-ClickThrough $hl
+  } elseif ($hl.Bounds.X -ne $bounds.X -or $hl.Bounds.Y -ne $bounds.Y -or $hl.Bounds.Width -ne $bounds.Width -or $hl.Bounds.Height -ne $bounds.Height) {
+    # 不重设隐藏；同一元素内移动时不重复改窗口，避免 WinForms 分层窗体重绘闪烁
+    $hl.Bounds = $bounds
+  }
 }
 function Hide-Highlight { $hl.Hide() }
 function Get-DeepestAtPoint([double]$px, [double]$py) {
@@ -187,6 +192,19 @@ $prevSpace = $false
 $prevPgUp = $false
 $prevPgDn = $false
 $prevF2 = $false
+$cacheX = [int]::MinValue
+$cacheY = [int]::MinValue
+$cacheEl = $null
+$cacheCtrlName = ""
+$cacheName = ""
+$cachePaneX = 0
+$cachePaneY = 0
+$cachePaneW = 0
+$cachePaneH = 0
+$cacheAt = [DateTime]::MinValue
+$lastQueryAt = [DateTime]::MinValue
+$queryGapMs = 45
+$cacheTtlMs = 1800
 Write-PickLog "loop start"
 while ($true) {
   try {
@@ -199,17 +217,49 @@ while ($true) {
     if ($stage -eq "hover") {
       $ptX = [System.Windows.Forms.Cursor]::Position.X
       $ptY = [System.Windows.Forms.Cursor]::Position.Y
-      $el = Get-DeepestAtPoint $ptX $ptY
+      $now = [DateTime]::UtcNow
+      $samePoint = ($ptX -eq $cacheX -and $ptY -eq $cacheY)
+      $insideCache = $cacheEl -ne $null -and $ptX -ge $cachePaneX -and $ptX -lt ($cachePaneX + $cachePaneW) -and $ptY -ge $cachePaneY -and $ptY -lt ($cachePaneY + $cachePaneH)
+      $cacheAlive = ($null -ne $cacheEl -and ($now - $cacheAt).TotalMilliseconds -le $cacheTtlMs)
+      # 鼠标还在当前元素矩形内部时不丢弃缓存；UIA 查询完成后才切换，避免移动过程先隐藏再显示
+      if (-not $samePoint -and -not $insideCache) { $cacheEl = $null }
+      if (-not $cacheAlive -and ($now - $lastQueryAt).TotalMilliseconds -ge $queryGapMs) {
+        $queryX = $ptX
+        $queryY = $ptY
+        $lastQueryAt = [DateTime]::UtcNow
+        try { $queried = Get-DeepestAtPoint $queryX $queryY } catch { $queried = $null; Write-PickLog ("QUERY ERR: " + $_.Exception.Message) }
+        $afterX = [System.Windows.Forms.Cursor]::Position.X
+        $afterY = [System.Windows.Forms.Cursor]::Position.Y
+        if ($afterX -ne $queryX -or $afterY -ne $queryY) {
+          $queried = $null
+          Write-PickLog "query stale: cursor moved"
+        }
+        if ($null -ne $queried) {
+          $qc = $queried.Current
+          $qr = $qc.BoundingRectangle
+          $cacheEl = $queried
+          $cacheX = $queryX
+          $cacheY = $queryY
+          $cacheAt = [DateTime]::UtcNow
+          $cacheCtrlName = $qc.ControlType.ProgrammaticName.Replace("ControlType.", "")
+          $cacheName = [string]$qc.Name
+          $cachePaneX = [int]$qr.X
+          $cachePaneY = [int]$qr.Y
+          $cachePaneW = [int]$qr.Width
+          $cachePaneH = [int]$qr.Height
+        } else {
+          $cacheEl = $null
+          $cacheCtrlName = ""
+          $cacheName = ""
+        }
+      }
+      $el = $cacheEl
       if ($el) {
-        $c = $el.Current
-        $ctrlName = $c.ControlType.ProgrammaticName.Replace("ControlType.", "")
-        $r = $c.BoundingRectangle
-        $paneX = [int]$r.X; $paneY = [int]$r.Y; $paneW = [int]$r.Width; $paneH = [int]$r.Height
-        Set-HighlightBox $paneX $paneY $paneW $paneH
-        $aiHint = if ($ctrlName -eq "Pane") { "   [F2]AI识别" } else { "" }
-        $infoLabel.Text = "名称: " + $c.Name + "   类型: " + $ctrlName + "   [空格]锁定" + $aiHint + "   [Esc]取消"
+        Set-HighlightBox $cachePaneX $cachePaneY $cachePaneW $cachePaneH
+        $aiHint = if ($cacheCtrlName -eq "Pane") { "   [F2]AI识别" } else { "" }
+        $infoLabel.Text = "名称: " + $cacheName + "   类型: " + $cacheCtrlName + "   [空格]锁定" + $aiHint + "   [Esc]取消"
       } else { Hide-Highlight }
-      if ($f2 -and -not $prevF2 -and $el -and $ctrlName -eq "Pane") {
+      if ($f2 -and -not $prevF2 -and $el -and $cacheCtrlName -eq "Pane") {
         Write-PickLog "f2 ai"
         $walker = [System.Windows.Automation.TreeWalker]::ControlViewWalker
         $cur = $el
@@ -218,7 +268,7 @@ while ($true) {
           $topTitle = [string]$cur.Current.Name
           $cur = $walker.GetParent($cur)
         }
-        $result = @{ aiRequested = $true; cursorX = $ptX; cursorY = $ptY; window = $topTitle; name = [string]$c.Name; controlType = $ctrlName; paneX = $paneX; paneY = $paneY; paneW = $paneW; paneH = $paneH }
+        $result = @{ aiRequested = $true; cursorX = $ptX; cursorY = $ptY; window = $topTitle; name = $cacheName; controlType = $cacheCtrlName; paneX = $cachePaneX; paneY = $cachePaneY; paneW = $cachePaneW; paneH = $cachePaneH }
         break
       }
       if ($space -and -not $prevSpace -and $el) {
@@ -226,7 +276,7 @@ while ($true) {
         $stage = "locked"
         $chain = New-Object System.Collections.ArrayList
         $walker = [System.Windows.Automation.TreeWalker]::ControlViewWalker
-        $cur = $el
+        $cur = $cacheEl
         while ($cur -and -not $cur.Equals([System.Windows.Automation.AutomationElement]::RootElement)) {
           $c = $cur.Current
           $r = $c.BoundingRectangle
diff --git a/resources/models/ui-icon-detect/model.onnx b/resources/models/ui-icon-detect/model.onnx
new file mode 100644
index 0000000..9189fde
Binary files /dev/null and b/resources/models/ui-icon-detect/model.onnx differ
diff --git a/resources/models/ui-icon-detect/preprocessor_config.json b/resources/models/ui-icon-detect/preprocessor_config.json
new file mode 100644
index 0000000..4ce8c7b
--- /dev/null
+++ b/resources/models/ui-icon-detect/preprocessor_config.json
@@ -0,0 +1,12 @@
+{
+  "do_normalize": false,
+  "do_rescale": true,
+  "do_resize": true,
+  "feature_extractor_type": "ImageFeatureExtractor",
+  "resample": 2,
+  "rescale_factor": 0.00392156862745098,
+  "size_divisor": 16,
+  "size": {
+    "longest_edge": 640
+  }
+}
\ No newline at end of file
diff --git a/scripts/package.mjs b/scripts/package.mjs
index 438166d..d3c9979 100644
--- a/scripts/package.mjs
+++ b/scripts/package.mjs
@@ -115,9 +115,13 @@ if (!fs.existsSync(path.join(builderUnpacked, "RPA Studio.exe"))) {
 }
 const packagedCatalogPath = path.join(builderUnpacked, "resources", "catalog.json");
 const packagedFixturePath = path.join(builderUnpacked, "resources", "fixture", "index.html");
+const packagedPickerPath = path.join(builderUnpacked, "resources", "desktop-picker.ps1");
 if (!fs.existsSync(packagedCatalogPath) || !fs.existsSync(packagedFixturePath)) {
   throw new Error("打包产物缺少运行资源：resources/catalog.json 或 resources/fixture/index.html");
 }
+if (!fs.existsSync(packagedPickerPath)) {
+  throw new Error("打包产物缺少桌面拾取脚本：resources/desktop-picker.ps1 未随包发布");
+}
 const packagedCatalog = JSON.parse(fs.readFileSync(packagedCatalogPath, "utf8"));
 if (!Array.isArray(packagedCatalog.components) || packagedCatalog.components.length === 0) {
   throw new Error("打包产物 resources/catalog.json 无效：组件目录为空");
@@ -216,6 +220,9 @@ if (!fs.existsSync(path.join(installSmokeTarget, "RPA Studio.exe"))) {
 if (!fs.existsSync(path.join(installSmokeTarget, "resources", "catalog.json"))) {
   throw new Error("NSIS 安装包冒烟失败：catalog.json 未安装");
 }
+if (!fs.existsSync(path.join(installSmokeTarget, "resources", "desktop-picker.ps1"))) {
+  throw new Error("NSIS 安装包冒烟失败：desktop-picker.ps1 未安装");
+}
 await verifyLaunch(path.join(installSmokeTarget, "RPA Studio.exe"), "installed");
 removeTree(installSmokeRoot);
 
diff --git a/src/main/browser/session.ts b/src/main/browser/session.ts
index eabc071..5c383f1 100644
--- a/src/main/browser/session.ts
+++ b/src/main/browser/session.ts
@@ -233,6 +233,18 @@ export class BrowserSession implements BrowserLike {
             return !name || accessibleName.includes(name);
           }) || null;
         }
+        if (spec.kind === 'css') {
+          const segments = String(spec.value).split('>>>').map((item) => item.trim()).filter(Boolean);
+          let root = document;
+          for (let index = 0; index < segments.length; index += 1) {
+            const found = root.querySelector(segments[index]);
+            if (!found) return null;
+            if (index === segments.length - 1) return found;
+            root = found.shadowRoot;
+            if (!root) return null;
+          }
+          return null;
+        }
         return document.querySelector(spec.value);
       };
       const el = find();
@@ -269,33 +281,46 @@ export class BrowserSession implements BrowserLike {
         try { return document.querySelectorAll(candidate).length === 1; } catch { return false; }
       };
       const cssPath = (node) => {
-        if (node.id) {
-          const candidate = '#' + cssEscape(node.id);
-          if (isUnique(candidate)) return candidate;
-        }
-        const testId = node.getAttribute('data-testid');
-        if (testId) {
-          const candidate = '[data-testid="' + cssEscape(testId) + '"]';
-          if (isUnique(candidate)) return candidate;
-        }
-        const parts = [];
+        const localPath = (start, root) => {
+          const stable = (item) => {
+            if (item.id) {
+              const candidate = '#' + cssEscape(item.id);
+              try { if (root.querySelectorAll(candidate).length === 1) return candidate; } catch {}
+            }
+            const testId = item.getAttribute('data-testid');
+            if (testId) {
+              const candidate = '[data-testid="' + cssEscape(testId) + '"]';
+              try { if (root.querySelectorAll(candidate).length === 1) return candidate; } catch {}
+            }
+            return null;
+          };
+          const parts = [];
+          let current = start;
+          while (current && current.nodeType === 1) {
+            const stableCandidate = stable(current);
+            if (stableCandidate) {
+              parts.unshift(stableCandidate);
+              return parts.join(' > ');
+            }
+            let part = current.tagName.toLowerCase();
+            const parent = current.parentElement;
+            if (parent) {
+              const siblings = [...parent.children].filter((item) => item.tagName === current.tagName);
+              if (siblings.length > 1) part += ':nth-of-type(' + (siblings.indexOf(current) + 1) + ')';
+            }
+            parts.unshift(part);
+            current = parent;
+          }
+          return parts.join(' > ');
+        };
+        const trees = [];
         let current = node;
         while (current && current.nodeType === 1) {
-          let part = current.tagName.toLowerCase();
-          if (current.id) {
-            part += '#' + cssEscape(current.id);
-            parts.unshift(part);
-            break;
-          }
-          const parent = current.parentElement;
-          if (parent) {
-            const siblings = [...parent.children].filter((item) => item.tagName === current.tagName);
-            if (siblings.length > 1) part += ':nth-of-type(' + (siblings.indexOf(current) + 1) + ')';
-          }
-          parts.unshift(part);
-          current = parent;
+          const root = current.getRootNode();
+          trees.unshift(localPath(current, root));
+          current = root instanceof ShadowRoot ? root.host : null;
         }
-        return parts.join(' > ');
+        return trees.join(' >>> ');
       };
       const describe = (node) => ({
         kind: 'css',
@@ -1143,33 +1168,46 @@ export class BrowserSession implements BrowserLike {
         try { return document.querySelectorAll(candidate).length === 1; } catch { return false; }
       };
       const cssPath = (node) => {
-        if (node.id) {
-          const candidate = '#' + cssEscape(node.id);
-          if (isUnique(candidate)) return candidate;
-        }
-        const testId = node.getAttribute('data-testid');
-        if (testId) {
-          const candidate = '[data-testid="' + cssEscape(testId) + '"]';
-          if (isUnique(candidate)) return candidate;
-        }
-        const parts = [];
+        const localPath = (start, root) => {
+          const stable = (item) => {
+            if (item.id) {
+              const candidate = '#' + cssEscape(item.id);
+              try { if (root.querySelectorAll(candidate).length === 1) return candidate; } catch {}
+            }
+            const testId = item.getAttribute('data-testid');
+            if (testId) {
+              const candidate = '[data-testid="' + cssEscape(testId) + '"]';
+              try { if (root.querySelectorAll(candidate).length === 1) return candidate; } catch {}
+            }
+            return null;
+          };
+          const parts = [];
+          let current = start;
+          while (current && current.nodeType === 1) {
+            const stableCandidate = stable(current);
+            if (stableCandidate) {
+              parts.unshift(stableCandidate);
+              return parts.join(' > ');
+            }
+            let part = current.tagName.toLowerCase();
+            const parent = current.parentElement;
+            if (parent) {
+              const siblings = [...parent.children].filter((item) => item.tagName === current.tagName);
+              if (siblings.length > 1) part += ':nth-of-type(' + (siblings.indexOf(current) + 1) + ')';
+            }
+            parts.unshift(part);
+            current = parent;
+          }
+          return parts.join(' > ');
+        };
+        const trees = [];
         let current = node;
         while (current && current.nodeType === 1) {
-          let part = current.tagName.toLowerCase();
-          if (current.id) {
-            part += '#' + cssEscape(current.id);
-            parts.unshift(part);
-            break;
-          }
-          const parent = current.parentElement;
-          if (parent) {
-            const siblings = [...parent.children].filter((item) => item.tagName === current.tagName);
-            if (siblings.length > 1) part += ':nth-of-type(' + (siblings.indexOf(current) + 1) + ')';
-          }
-          parts.unshift(part);
-          current = parent;
+          const root = current.getRootNode();
+          trees.unshift(localPath(current, root));
+          current = root instanceof ShadowRoot ? root.host : null;
         }
-        return parts.join(' > ');
+        return trees.join(' >>> ');
       };
       const describe = (node) => {
         const rect = node.getBoundingClientRect();
@@ -1200,9 +1238,33 @@ export class BrowserSession implements BrowserLike {
         badge.style.left = Math.max(0, rect.left) + 'px';
         badge.style.top = Math.max(0, rect.top - 30) + 'px';
       };
+      const interactiveSelector = [
+        'button', 'a', 'input', 'select', 'textarea', 'summary', 'label', 'option',
+        '[role]', '[onclick]', '[contenteditable]', '[tabindex]'
+      ].join(',');
+      const closestInteractive = (start) => {
+        let node = start;
+        while (node && node.nodeType === 1) {
+          try { if (node.matches(interactiveSelector)) return node; } catch {}
+          if (node.parentElement) {
+            node = node.parentElement;
+            continue;
+          }
+          const root = node.getRootNode();
+          node = root instanceof ShadowRoot ? root.host : null;
+        }
+        return start;
+      };
       const firstElement = (event) => {
-        for (const item of event.composedPath()) if (item instanceof Element) return item;
-        return event.target instanceof Element ? event.target : null;
+        let deepest = null;
+        for (const item of event.composedPath()) {
+          if (item instanceof Element) {
+            deepest = item;
+            break;
+          }
+        }
+        if (!deepest && event.target instanceof Element) deepest = event.target;
+        return closestInteractive(deepest);
       };
       const move = (event) => {
         if (finished) return;
diff --git a/src/main/capabilities/ai-service.ts b/src/main/capabilities/ai-service.ts
index 2dcb5a6..e859c11 100644
--- a/src/main/capabilities/ai-service.ts
+++ b/src/main/capabilities/ai-service.ts
@@ -34,16 +34,29 @@ export function requireAiConfig(options: { vision?: boolean } = {}): ResolvedAiC
   return config;
 }
 
-export async function chatCompletion(messages: ChatMessage[], options: { temperature?: number; vision?: boolean } = {}): Promise<string> {
+export interface ChatCompletionOptions {
+  temperature?: number;
+  vision?: boolean;
+  /** 交互式请求（用户正在等待，如 AI 拾取）：默认短超时 45s、最多 2 次，超时不重试 */
+  interactive?: boolean;
+  /** 单次请求超时毫秒数；显式指定后覆盖交互式/后台默认值 */
+  timeoutMs?: number;
+  /** 最大尝试次数；显式指定后覆盖交互式/后台默认值 */
+  maxAttempts?: number;
+}
+
+export async function chatCompletion(messages: ChatMessage[], options: ChatCompletionOptions = {}): Promise<string> {
   const config = requireAiConfig({ vision: options.vision });
   const base = config.baseUrl.replace(/\/+$/, "");
-  const maxAttempts = 3;
+  const interactive = options.interactive === true;
+  const timeoutMs = Math.max(1000, Math.round(options.timeoutMs ?? (interactive ? 45000 : 170000)));
+  const maxAttempts = Math.max(1, Math.round(options.maxAttempts ?? (interactive ? 2 : 3)));
 
   for (let attempt = 1; attempt <= maxAttempts; attempt++) {
     let response: Response;
     try {
       response = await fetch(base + "/chat/completions", {
-        signal: AbortSignal.timeout(170000),
+        signal: AbortSignal.timeout(timeoutMs),
         method: "POST",
         headers: { "Content-Type": "application/json", Authorization: "Bearer " + config.apiKey },
         body: JSON.stringify({
@@ -53,6 +66,10 @@ export async function chatCompletion(messages: ChatMessage[], options: { tempera
         })
       });
     } catch (error) {
+      // 超时（AbortSignal.timeout）不重试：重发只会成倍拉长等待
+      if ((error as { name?: string }).name === "TimeoutError") {
+        throw new RuntimeError("ai_request_timeout", `AI 请求超时（${timeoutMs}ms），请检查网络或稍后重试`);
+      }
       if (attempt === maxAttempts) throw error;
       await new Promise(resolve => setTimeout(resolve, 1200 * attempt));
       continue;
diff --git a/src/main/capabilities/cv-vision.test.ts b/src/main/capabilities/cv-vision.test.ts
new file mode 100644
index 0000000..f5ece19
--- /dev/null
+++ b/src/main/capabilities/cv-vision.test.ts
@@ -0,0 +1,50 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+const powershellMock = vi.hoisted(() => vi.fn(async () => '{"ok":true}'));
+const detectMock = vi.hoisted(() => vi.fn(async () => ({
+  texts: [{ text: "确认保存", box: [[0, 0], [20, 0], [20, 10], [0, 10]] }]
+})));
+
+vi.mock("./desktop", () => ({ powershell: powershellMock }));
+vi.mock("@gutenye/ocr-node", () => ({
+  default: { create: vi.fn(async () => ({ detect: detectMock })) }
+}));
+
+import { getHandler } from "../engine/registry";
+import "./cv-vision";
+
+describe("CV vision capabilities", () => {
+  beforeEach(() => {
+    powershellMock.mockClear();
+    detectMock.mockClear();
+  });
+
+  it.each(["CV_GetText", "CV_SetText", "CV_MouseHover", "CV_MouseClick"])("registers %s", (id) => {
+    expect(getHandler(id), id).toBeTruthy();
+  });
+
+  it("rejects an empty CV selector", async () => {
+    await expect(getHandler("CV_GetText")!({ log: vi.fn() } as any, { id: "n1", componentId: "CV_GetText" } as any, {}, {} as any))
+      .rejects.toThrow("CV 目标为空");
+  });
+
+  it("locates OCR text and clicks its center with offset", async () => {
+    const handler = getHandler("CV_MouseClick")!;
+    await expect(handler({ log: vi.fn() } as any, { id: "n1", componentId: "CV_MouseClick" } as any, {
+      cvSelector: "确认",
+      offsetX: 5,
+      offsetY: 2
+    }, {} as any)).resolves.toBe(true);
+    expect(detectMock).toHaveBeenCalledTimes(1);
+    const script = String((powershellMock.mock.calls as unknown as Array<[string]>)[1]?.[0] ?? "");
+    expect(script).toContain("SetCursorPos(15, 7)");
+    expect(script).toContain("mouse_event(0x0002");
+  });
+
+  it("returns matched OCR text", async () => {
+    await expect(getHandler("CV_GetText")!({ log: vi.fn() } as any, { id: "n2", componentId: "CV_GetText" } as any, {
+      cvSelector: "保存"
+    }, {} as any)).resolves.toBe("确认保存");
+    expect(powershellMock).toHaveBeenCalledTimes(1);
+  });
+});
diff --git a/src/main/capabilities/cv-vision.ts b/src/main/capabilities/cv-vision.ts
index 88222ac..d3935ee 100644
--- a/src/main/capabilities/cv-vision.ts
+++ b/src/main/capabilities/cv-vision.ts
@@ -1,5 +1,5 @@
 import { register } from "../engine/registry";
-import { RuntimeError } from "../engine/types";
+import { RuntimeError, type CapabilityHandler } from "../engine/types";
 import { powershell } from "./desktop";
 import fs from "node:fs";
 import os from "node:os";
@@ -57,4 +57,95 @@ export async function locateTextOnScreen(target: string, nodeId: string): Promis
   const cx = (Math.min(...xs) + Math.max(...xs)) / 2;
   const cy = (Math.min(...ys) + Math.max(...ys)) / 2;
   return { target: keyword, text: hit.text, cx, cy };
-}
\ No newline at end of file
+}
+function screenUser32(): string {
+  return [
+    "Add-Type -MemberDefinition '[DllImport(\"user32.dll\")] public static extern bool SetProcessDpiAwarenessContext(IntPtr value); [DllImport(\"user32.dll\")] public static extern bool SetCursorPos(int X, int Y); [DllImport(\"user32.dll\")] public static extern void mouse_event(uint flags, uint dx, uint dy, uint data, UIntPtr extra);' -Name RpaCvInput -Namespace RpaCv | Out-Null;",
+    "try { [RpaCv.RpaCvInput]::SetProcessDpiAwarenessContext([IntPtr](-4)) | Out-Null } catch {}"
+  ].join("\n");
+}
+
+function coordinateScript(x: number, y: number, action: "hover" | "click" | "doubleClick" | "paste", value = ""): string {
+  const lines = [
+    screenUser32(),
+    "if (-not [RpaCv.RpaCvInput]::SetCursorPos(" + x + ", " + y + ")) { throw 'cursor_move_failed' }",
+    "Start-Sleep -Milliseconds 100"
+  ];
+  if (action !== "hover") {
+    lines.push(
+      "[RpaCv.RpaCvInput]::mouse_event(0x0002,0,0,0,[UIntPtr]::Zero)",
+      "Start-Sleep -Milliseconds 40",
+      "[RpaCv.RpaCvInput]::mouse_event(0x0004,0,0,0,[UIntPtr]::Zero)"
+    );
+  }
+  if (action === "doubleClick") {
+    lines.push(
+      "Start-Sleep -Milliseconds 70",
+      "[RpaCv.RpaCvInput]::mouse_event(0x0002,0,0,0,[UIntPtr]::Zero)",
+      "Start-Sleep -Milliseconds 40",
+      "[RpaCv.RpaCvInput]::mouse_event(0x0004,0,0,0,[UIntPtr]::Zero)"
+    );
+  }
+  if (action === "paste") {
+    lines.push(
+      "Add-Type -AssemblyName System.Windows.Forms",
+      "Start-Sleep -Milliseconds 220",
+      "$saved = $null",
+      "try { $saved = [System.Windows.Forms.Clipboard]::GetText() } catch {}",
+      "try { [System.Windows.Forms.Clipboard]::SetText(" + JSON.stringify(value) + ") } catch { throw 'clipboard_failed' }",
+      "Start-Sleep -Milliseconds 80",
+      "[System.Windows.Forms.SendKeys]::SendWait('^a')",
+      "Start-Sleep -Milliseconds 60",
+      "[System.Windows.Forms.SendKeys]::SendWait('^v')",
+      "Start-Sleep -Milliseconds 280",
+      "try { if ($saved) { [System.Windows.Forms.Clipboard]::SetText($saved) } else { [System.Windows.Forms.Clipboard]::Clear() } } catch {}"
+    );
+  }
+  lines.push("Write-Output (@{ ok = $true; x = " + x + "; y = " + y + " } | ConvertTo-Json -Compress)");
+  return lines.join("\n");
+}
+
+async function locateFor(args: Record<string, unknown>, node: { id: string; componentId: string }) {
+  const target = cvTarget(args, node.id);
+  const located = await locateTextOnScreen(target, node.id);
+  const offsetX = Number(args.offsetX ?? args.offset_x ?? 0) || 0;
+  const offsetY = Number(args.offsetY ?? args.offset_y ?? 0) || 0;
+  return { located, x: Math.round(located.cx + offsetX), y: Math.round(located.cy + offsetY) };
+}
+
+const getTextGround: CapabilityHandler = async (ctx, node, args) => {
+  const target = cvTarget(args, node.id);
+  const result = await locateTextOnScreen(target, node.id);
+  ctx.log("CV_GetText found: " + result.text, "info", node.id);
+  return result.text;
+};
+
+const setTextGround: CapabilityHandler = async (ctx, node, args) => {
+  const value = String(args.value ?? args.text ?? "");
+  if (!value) throw new RuntimeError("validation_error", "CV 文本输入内容为空", { nodeId: node.id, componentId: node.componentId });
+  const { x, y } = await locateFor(args, node);
+  await powershell(coordinateScript(x, y, "paste", value), 30000);
+  ctx.log("CV_SetText input at " + x + "," + y, "info", node.id);
+  return true;
+};
+
+const mouseHoverGround: CapabilityHandler = async (ctx, node, args) => {
+  const { x, y } = await locateFor(args, node);
+  await powershell(coordinateScript(x, y, "hover"), 15000);
+  ctx.log("CV_MouseHover hover at " + x + "," + y, "info", node.id);
+  return true;
+};
+
+const mouseClickGround: CapabilityHandler = async (ctx, node, args) => {
+  const { x, y } = await locateFor(args, node);
+  const clickway = String(args.clickway ?? "0");
+  const action = clickway === "1" || /double/i.test(clickway) ? "doubleClick" as const : "click" as const;
+  await powershell(coordinateScript(x, y, action), 15000);
+  ctx.log("CV_MouseClick " + action + " at " + x + "," + y, "info", node.id);
+  return true;
+};
+
+register("CV_GetText", getTextGround);
+register("CV_SetText", setTextGround);
+register("CV_MouseHover", mouseHoverGround);
+register("CV_MouseClick", mouseClickGround);
diff --git a/src/main/capabilities/desktop-uia.test.ts b/src/main/capabilities/desktop-uia.test.ts
index c83024b..c4c27a7 100644
--- a/src/main/capabilities/desktop-uia.test.ts
+++ b/src/main/capabilities/desktop-uia.test.ts
@@ -1,7 +1,16 @@
-import { describe, expect, it } from "vitest";
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+const powershellMock = vi.hoisted(() => vi.fn(async () => '{"clicked":true,"value":"cached"}'));
+
+vi.mock("./desktop", () => ({
+  powershell: powershellMock,
+  psString: (value: unknown) => JSON.stringify(String(value ?? ""))
+}));
+
 import { getHandler } from "../engine/registry";
 import { parseUiaSelector } from "./desktop-uia";
 
+
 describe("desktop UIA ground capabilities", () => {
   it.each([
     "Element_ClickGround",
@@ -21,7 +30,10 @@ describe("desktop UIA ground capabilities", () => {
     "Element_MouseToGround",
     "Element_ScrollIntoViewGround",
     "Element_ChildrenGround",
-    "Element_DesktopChildren"
+    "Element_DesktopChildren",
+    "Element_RestoreWindowGround",
+    "Element_ActivateWindowGround",
+    "Element_MoveWindowGround"
   ])("registers %s", (id) => {
     expect(getHandler(id), id).toBeTruthy();
   });
@@ -57,3 +69,61 @@ describe("desktop UIA ground capabilities", () => {
     expect(() => parseUiaSelector("screen=1,2,3")).toThrow();
   });
 });
+
+describe("desktop UIA selector compatibility", () => {
+  beforeEach(() => {
+    powershellMock.mockClear();
+    powershellMock.mockImplementation(async () => '{"clicked":true,"value":"cached"}');
+  });
+
+  it("maps compound .NET ControlType names explicitly and preserves raw input", () => {
+    expect(parseUiaSelector("type=combobox")).toMatchObject({ type: "ComboBox", typeRaw: "combobox" });
+    expect(parseUiaSelector("type=listitem")).toMatchObject({ type: "ListItem", typeRaw: "listitem" });
+    expect(parseUiaSelector("type=datagrid")).toMatchObject({ type: "DataGrid", typeRaw: "datagrid" });
+  });
+
+  it("parses a stable ancestor path", () => {
+    const selector = parseUiaSelector("window=App;path=Pane>Edit[name=用户名;index=2]>Button[name=确定]");
+    expect(selector.path).toEqual([
+      { type: "Pane" },
+      { type: "Edit", name: "用户名", index: 2 },
+      { type: "Button", name: "确定" }
+    ]);
+  });
+
+  const context = { log: vi.fn() } as any;
+  const node = { id: "n1", componentId: "Element_ClickGround" } as any;
+
+  it("dispatches screen selectors to coordinate execution instead of UIA lookup", async () => {
+    const args = {
+      selector: "screen=100,200,320,80;ai=1",
+      clickway: "1"
+    };
+    await expect(getHandler("Element_ClickGround")!(context, node, args, {} as any)).resolves.toEqual({ clicked: true, value: "cached" });
+    const script = String((powershellMock.mock.calls as unknown as Array<[string]>)[0]?.[0] ?? "");
+    expect(script).toContain("RpaScreenInput");
+    expect(script).toContain("SetCursorPos(260, 240)");
+    expect(script).not.toContain("Find-RpaWindow");
+  });
+
+  it("dispatches screen setValue to coordinate paste", async () => {
+    await getHandler("Element_SetValueGround")!(context, node, {
+      selector: "screen=10,20,30,40",
+      value: "hello"
+    }, {} as any);
+    const script = String((powershellMock.mock.calls as unknown as Array<[string]>)[0]?.[0] ?? "");
+    expect(script).toContain("screen_coordinate_paste");
+    expect(script).toContain("hello");
+    expect(script).not.toContain("Find-RpaWindow");
+  });
+
+  it("treats a picked screen rectangle as present for exists/wait", async () => {
+    await expect(getHandler("Element_ExistsGround")!(context, node, {
+      selector: "screen=1,2,3,4"
+    }, {} as any)).resolves.toBe(true);
+    await expect(getHandler("Element_WaitElementGround")!(context, node, {
+      selector: "screen=1,2,3,4"
+    }, {} as any)).resolves.toBe(true);
+    expect(powershellMock).not.toHaveBeenCalled();
+  });
+});
diff --git a/src/main/capabilities/desktop-uia.ts b/src/main/capabilities/desktop-uia.ts
index e02436a..9ed3f97 100644
--- a/src/main/capabilities/desktop-uia.ts
+++ b/src/main/capabilities/desktop-uia.ts
@@ -8,12 +8,124 @@ interface UiaSelector {
   window?: string;
   name?: string;
   id?: string;
+  /** .NET ControlType 字段名（已映射，如 ComboBox/ListItem/DataGrid） */
   type?: string;
+  /** 用户输入的原始类型字符串 */
+  typeRaw?: string;
   className?: string;
   screen?: { x: number; y: number; w: number; h: number };
+  /** 从窗口（或桌面根）到目标元素的祖先路径，逐段用稳定属性定位 */
+  path?: Array<{ type?: string; name?: string; id?: string; className?: string; index?: number }>;
   index: number;
 }
 
+/** ControlType 显式映射：type= 值 → .NET ControlType 字段名；复合名称不得用首字母大写推导 */
+const UIA_CONTROL_TYPE_MAP: Record<string, string> = {
+  button: "Button",
+  calendar: "Calendar",
+  checkbox: "CheckBox",
+  combo: "ComboBox",
+  combobox: "ComboBox",
+  custom: "Custom",
+  datagrid: "DataGrid",
+  dataitem: "DataItem",
+  document: "Document",
+  edit: "Edit",
+  group: "Group",
+  header: "Header",
+  headeritem: "HeaderItem",
+  hyperlink: "Hyperlink",
+  image: "Image",
+  list: "List",
+  listbox: "List",
+  listitem: "ListItem",
+  menu: "Menu",
+  menubar: "MenuBar",
+  menuitem: "MenuItem",
+  pane: "Pane",
+  progressbar: "ProgressBar",
+  radio: "RadioButton",
+  radiobutton: "RadioButton",
+  scrollbar: "ScrollBar",
+  separator: "Separator",
+  slider: "Slider",
+  spinner: "Spinner",
+  splitbutton: "SplitButton",
+  statusbar: "StatusBar",
+  tab: "Tab",
+  tabitem: "TabItem",
+  table: "Table",
+  text: "Text",
+  thumb: "Thumb",
+  titlebar: "TitleBar",
+  toolbar: "ToolBar",
+  tooltip: "ToolTip",
+  tree: "Tree",
+  treeitem: "TreeItem",
+  window: "Window"
+};
+
+function normalizeUiaControlType(raw: string): string {
+  const key = raw.trim().toLowerCase().replace(/[\s_-]+/g, "");
+  return UIA_CONTROL_TYPE_MAP[key] ?? raw.charAt(0).toUpperCase() + raw.slice(1).toLowerCase();
+}
+
+/** 解析 path= 段：Pane>Edit[name=用户名;index=2]，段间用 > 分隔，段内 [prop=value;...] */
+function parseUiaSelectorPath(value: string): UiaSelector["path"] {
+  // 按括号外的 > 分段，避免 path=Edit[name=a>b] 这类属性值被错误拆开
+  const rawSegments: string[] = [];
+  let buffer = "";
+  let bracketDepth = 0;
+  for (const char of value) {
+    if (char === "[") bracketDepth += 1;
+    else if (char === "]") bracketDepth = Math.max(0, bracketDepth - 1);
+    if (char === ">" && bracketDepth === 0) {
+      rawSegments.push(buffer);
+      buffer = "";
+    } else {
+      buffer += char;
+    }
+  }
+  rawSegments.push(buffer);
+  const segments = rawSegments.map(item => item.trim()).filter(Boolean);
+  if (!segments.length) return undefined;
+  const path: NonNullable<UiaSelector["path"]> = [];
+  for (const segment of segments) {
+    const match = segment.match(/^([A-Za-z][A-Za-z0-9_]*)?\s*(?:\[(.*)\])?$/);
+    if (!match) {
+      throw new RuntimeError("validation_error", `UIA 路径段无效：${segment}`, { recommendation: "使用 Type[name=xx;id=yy;class=zz;index=1] > 下一级" });
+    }
+    const entry: NonNullable<UiaSelector["path"]>[number] = {};
+    if (match[1]) entry.type = normalizeUiaControlType(match[1]);
+    const props = match[2];
+    if (props) {
+      for (const pair of props.split(/[;，,]/).map(item => item.trim()).filter(Boolean)) {
+        const eq = pair.indexOf("=");
+        if (eq <= 0) {
+          throw new RuntimeError("validation_error", `UIA 路径属性无效：${pair}`, { recommendation: "使用 name= / id= / class= / index=" });
+        }
+        const key = pair.slice(0, eq).trim().toLowerCase();
+        const val = pair.slice(eq + 1).trim();
+        if (key === "name" || key === "n") entry.name = val;
+        else if (key === "id" || key === "automationid") entry.id = val;
+        else if (key === "class" || key === "classname") entry.className = val;
+        else if (key === "type" || key === "controltype") entry.type = normalizeUiaControlType(val);
+        else if (key === "index") {
+          const n = Number(val);
+          if (!Number.isFinite(n) || n < 1) throw new RuntimeError("validation_error", `UIA 路径 index 无效：${val}`);
+          entry.index = Math.floor(n);
+        }
+        else throw new RuntimeError("validation_error", `UIA 路径属性不支持：${key}`);
+      }
+    }
+    if (!entry.type && !entry.name && !entry.id && !entry.className) {
+      throw new RuntimeError("validation_error", `UIA 路径段缺少定位条件：${segment}`, { recommendation: "每段至少提供 type/name/id/class 之一" });
+    }
+    path.push(entry);
+  }
+  return path;
+}
+
 export function parseUiaSelector(raw: unknown): UiaSelector {
   let text = "";
   if (raw && typeof raw === "object") {
@@ -33,7 +145,22 @@ export function parseUiaSelector(raw: unknown): UiaSelector {
   }
   if (text.startsWith("uia:")) text = text.slice(4);
   const selector: UiaSelector = { index: 1 };
-  for (const part of text.split(";")) {
+  // path 段内的 [name=a;b] 不应被顶层 ; 拆开
+  const selectorParts: string[] = [];
+  let partBuffer = "";
+  let selectorBracketDepth = 0;
+  for (const char of text) {
+    if (char === "[") selectorBracketDepth += 1;
+    else if (char === "]") selectorBracketDepth = Math.max(0, selectorBracketDepth - 1);
+    if (char === ";" && selectorBracketDepth === 0) {
+      selectorParts.push(partBuffer);
+      partBuffer = "";
+    } else {
+      partBuffer += char;
+    }
+  }
+  selectorParts.push(partBuffer);
+  for (const part of selectorParts) {
     const eq = part.indexOf("=");
     if (eq <= 0) continue;
     const key = part.slice(0, eq).trim().toLowerCase();
@@ -42,7 +169,10 @@ export function parseUiaSelector(raw: unknown): UiaSelector {
     if (key === "window" || key === "win") selector.window = value;
     else if (key === "name" || key === "n") selector.name = value;
     else if (key === "id" || key === "automationid") selector.id = value;
-    else if (key === "type" || key === "controltype") selector.type = value.charAt(0).toUpperCase() + value.slice(1).toLowerCase();
+    else if (key === "type" || key === "controltype") {
+      selector.typeRaw = value;
+      selector.type = normalizeUiaControlType(value);
+    }
     else if (key === "class" || key === "classname") selector.className = value;
     else if (key === "screen") {
       const nums = value.split(/[,，]/).map(item => Number(item.trim()));
@@ -55,8 +185,11 @@ export function parseUiaSelector(raw: unknown): UiaSelector {
       const n = Number(value);
       if (Number.isFinite(n) && n >= 1) selector.index = Math.floor(n);
     }
+    else if (key === "path" || key === "ancestors" || key === "ancestorpath") {
+      selector.path = parseUiaSelectorPath(value);
+    }
   }
-  if (!selector.name && !selector.id && !selector.className && !selector.type && !selector.screen) {
+  if (!selector.name && !selector.id && !selector.className && !selector.type && !selector.screen && !selector.path) {
     throw new RuntimeError("validation_error", `桌面元素选择器缺少定位条件：${text}`, { recommendation: "至少提供 name= / id= / type= 之一" });
   }
   return selector;
@@ -69,6 +202,7 @@ const UIA_USING = [
 
 const FIND_HELPER = `
 function Find-RpaWindow([string]$winName) {
+  # 只读查找：绝不移动/还原/缩放窗口；还原、激活、移动是独立显式动作（Restore/Activate/Move-RpaWindow）
   $root = [System.Windows.Automation.AutomationElement]::RootElement
   if (-not $winName) { return $root }
   $cond = New-Object System.Windows.Automation.PropertyCondition([System.Windows.Automation.AutomationElement]::NameProperty, $winName)
@@ -78,41 +212,88 @@ function Find-RpaWindow([string]$winName) {
   $script:hitHwnd = [IntPtr]::Zero
   $cb = [WinRpa.RpaUser32Click+EnumProc]{ param($h, $l)
     $sb = New-Object System.Text.StringBuilder 256
-    [WinRpa.RpaUser32Click]::GetWindowTextW($h, $sb, 256) | Out-Null
+    [void][WinRpa.RpaUser32Click]::GetWindowTextW($h, $sb, 256)
     if ($sb.ToString() -eq $winName) { $script:hitHwnd = $h; return $false }
     return $true
   }
-  [WinRpa.RpaUser32Click]::EnumWindows($cb, [IntPtr]::Zero) | Out-Null
+  [void][WinRpa.RpaUser32Click]::EnumWindows($cb, [IntPtr]::Zero)
   if ($script:hitHwnd -ne [IntPtr]::Zero) {
-    [WinRpa.RpaUser32Click]::ShowWindow($script:hitHwnd, 9) | Out-Null
-    Start-Sleep -Milliseconds 500
-    $sw2 = [WinRpa.RpaUser32Click]::GetSystemMetrics(0)
-    $sh2 = [WinRpa.RpaUser32Click]::GetSystemMetrics(1)
-    $w2 = [Math]::Min(1400, $sw2 - 120)
-    $h2 = [Math]::Min(820, $sh2 - 120)
-    [WinRpa.RpaUser32Click]::SetWindowPos($script:hitHwnd, [IntPtr]::Zero, 60, 60, $w2, $h2, 0x0040) | Out-Null
-    Start-Sleep -Milliseconds 600
-    $win = $root.FindFirst([System.Windows.Automation.TreeScope]::Children, $cond)
-    if ($win) { return $win }
+    try { $win = [System.Windows.Automation.AutomationElement]::FromHandle($script:hitHwnd); if ($win) { return $win } } catch {}
   }
   throw "window_not_found:$winName"
 }
-function Find-RpaElement($parent, [string]$name, [string]$id, [string]$type, [string]$className, [int]$index) {
+function Restore-RpaWindow($win) {
+  # 显式动作：还原窗口（最小化/最大化状态恢复），不用于查询
+  Add-RpaUser32
+  $hwnd = [IntPtr]$win.Current.NativeWindowHandle
+  if ($hwnd -eq [IntPtr]::Zero) { throw "window_no_hwnd" }
+  [void][WinRpa.RpaUser32Click]::ShowWindow($hwnd, 9)
+  Start-Sleep -Milliseconds 300
+  return $hwnd
+}
+function Activate-RpaWindow($win) {
+  # 显式动作：把窗口带到前台
+  Add-RpaUser32
+  $hwnd = [IntPtr]$win.Current.NativeWindowHandle
+  if ($hwnd -eq [IntPtr]::Zero) { throw "window_no_hwnd" }
+  $fgHwnd = [WinRpa.RpaUser32Click]::GetForegroundWindow()
+  if ($fgHwnd -ne $hwnd) {
+    $fgPid = [uint32]0
+    $fgThread = [WinRpa.RpaUser32Click]::GetWindowThreadProcessId($fgHwnd, [ref]$fgPid)
+    $thisThread = [WinRpa.RpaUser32Click]::GetCurrentThreadId()
+    [void][WinRpa.RpaUser32Click]::AttachThreadInput($thisThread, $fgThread, $true)
+    [void][WinRpa.RpaUser32Click]::SetForegroundWindow($hwnd)
+    [void][WinRpa.RpaUser32Click]::AttachThreadInput($thisThread, $fgThread, $false)
+  }
+  Start-Sleep -Milliseconds 200
+  return $hwnd
+}
+function Move-RpaWindow($win, [int]$x, [int]$y, [int]$w, [int]$h) {
+  # 显式动作：移动/缩放窗口；w/h<=0 时使用安全默认尺寸
+  Add-RpaUser32
+  $hwnd = [IntPtr]$win.Current.NativeWindowHandle
+  if ($hwnd -eq [IntPtr]::Zero) { throw "window_no_hwnd" }
+  $sw = [WinRpa.RpaUser32Click]::GetSystemMetrics(0)
+  $sh = [WinRpa.RpaUser32Click]::GetSystemMetrics(1)
+  if ($w -le 0) { $w = [Math]::Min(1400, $sw - 120) }
+  if ($h -le 0) { $h = [Math]::Min(820, $sh - 120) }
+  [void][WinRpa.RpaUser32Click]::SetWindowPos($hwnd, [IntPtr]::Zero, $x, $y, $w, $h, 0x0040)
+  Start-Sleep -Milliseconds 300
+  return $hwnd
+}
+function Find-RpaElement($parent, [string]$name, [string]$id, [string]$type, [string]$className, [int]$index, [switch]$directChild) {
   $conds = @()
   if ($name) { $conds += New-Object System.Windows.Automation.PropertyCondition([System.Windows.Automation.AutomationElement]::NameProperty, $name) }
   if ($id) { $conds += New-Object System.Windows.Automation.PropertyCondition([System.Windows.Automation.AutomationElement]::AutomationIdProperty, $id) }
   if ($className) { $conds += New-Object System.Windows.Automation.PropertyCondition([System.Windows.Automation.AutomationElement]::ClassNameProperty, $className) }
   if ($type) {
-    try { $field = [System.Windows.Automation.ControlType].GetField($type); if ($field) { $conds += New-Object System.Windows.Automation.PropertyCondition([System.Windows.Automation.AutomationElement]::ControlTypeProperty, $field.GetValue($null)) } } catch {}
-    if ($ct) { $conds += New-Object System.Windows.Automation.PropertyCondition([System.Windows.Automation.AutomationElement]::ControlTypeProperty, $ct) }
+    # 不区分大小写匹配 .NET ControlType 字段，兼容 ComboBox/ListItem/DataGrid 等复合名称
+    try {
+      $ctField = [System.Windows.Automation.ControlType].GetFields([System.Reflection.BindingFlags]"Public,Static") | Where-Object { $_.Name -ieq $type } | Select-Object -First 1
+      if ($ctField) { $conds += New-Object System.Windows.Automation.PropertyCondition([System.Windows.Automation.AutomationElement]::ControlTypeProperty, $ctField.GetValue($null)) }
+    } catch {}
   }
   if ($conds.Count -eq 0) { throw "no_condition" }
   $final = if ($conds.Count -eq 1) { $conds[0] } else { New-Object System.Windows.Automation.AndCondition($conds) }
-  $all = $parent.FindAll([System.Windows.Automation.TreeScope]::Descendants, $final)
+  $scope = if ($directChild) { [System.Windows.Automation.TreeScope]::Children } else { [System.Windows.Automation.TreeScope]::Descendants }
+  $all = $parent.FindAll($scope, $final)
   if ($all.Count -eq 0) { return $null }
   $pick = [Math]::Min([Math]::Max($index, 1), $all.Count)
   return $all[$pick - 1]
 }
+function Find-RpaElementByPath($parent, [string]$pathJsonBase64) {
+  # 逐段走直接子级，用稳定属性（id/type/class 优先）解析窗口→目标祖先路径
+  $segments = [Text.Encoding]::UTF8.GetString([Convert]::FromBase64String($pathJsonBase64)) | ConvertFrom-Json
+  $cur = $parent
+  foreach ($seg in @($segments)) {
+    $ix = 1
+    try { if ($seg.index) { $ix = [int]$seg.index } } catch {}
+    $next = Find-RpaElement $cur $seg.name $seg.id $seg.type $seg.className $ix -directChild
+    if ($null -eq $next) { return $null }
+    $cur = $next
+  }
+  return $cur
+}
 function Add-RpaUser32 {
   Add-Type -MemberDefinition '[DllImport("user32.dll")] public static extern bool SetCursorPos(int X, int Y); [DllImport("user32.dll")] public static extern void mouse_event(uint flags, uint dx, uint dy, uint data, UIntPtr extra); [DllImport("user32.dll")] public static extern bool SetForegroundWindow(IntPtr hWnd); [DllImport("user32.dll")] public static extern IntPtr GetForegroundWindow(); [DllImport("user32.dll")] public static extern bool ShowWindow(IntPtr hWnd, int nCmdShow); [DllImport("user32.dll")] public static extern bool SetWindowPos(IntPtr hWnd, IntPtr after, int X, int Y, int cx, int cy, uint flags); [DllImport("user32.dll")] public static extern bool AttachThreadInput(uint idAttach, uint idAttachTo, bool fAttach); [DllImport("user32.dll")] public static extern uint GetWindowThreadProcessId(IntPtr hWnd, out uint processId); [DllImport("kernel32.dll")] public static extern uint GetCurrentThreadId(); [DllImport("user32.dll")] public static extern int GetSystemMetrics(int nIndex); [DllImport("user32.dll")] public static extern bool EnumWindows(EnumProc cb, IntPtr lp); public delegate bool EnumProc(IntPtr hWnd, IntPtr lp); [DllImport("user32.dll", CharSet = CharSet.Unicode)] public static extern int GetWindowTextW(IntPtr hWnd, System.Text.StringBuilder sb, int max);' -Name RpaUser32Click -Namespace Win32Rpa | Out-Null
 }
@@ -129,6 +310,21 @@ function target(args: Record<string, unknown>, node: { id: string; componentId:
   return { selector: parseUiaSelector(raw), args, node };
 }
 
+/** 生成“解析窗口 + 解析元素（路径优先，其次平铺属性）”的 PowerShell 语句，结果写入 $el */
+function elementResolveLines(s: UiaSelector): string {
+  const parentLine = "$parent = Find-RpaWindow " + psString(s.window ?? "");
+  if (s.path?.length) {
+    const pathB64 = Buffer.from(JSON.stringify(s.path), "utf8").toString("base64");
+    return parentLine + "\n$el = Find-RpaElementByPath $parent '" + pathB64 + "'";
+  }
+  return parentLine
+    + "\n$el = Find-RpaElement $parent " + psString(s.name ?? "")
+    + " " + psString(s.id ?? "")
+    + " " + psString(s.type ?? "")
+    + " " + psString(s.className ?? "")
+    + " " + s.index;
+}
+
 const SCREEN_USER32 = "Add-Type -MemberDefinition '[DllImport(\"user32.dll\")] public static extern bool SetProcessDpiAwarenessContext(IntPtr value); [DllImport(\"user32.dll\")] public static extern bool SetCursorPos(int X, int Y); [DllImport(\"user32.dll\")] public static extern void mouse_event(uint flags, uint dx, uint dy, uint data, UIntPtr extra);' -Name RpaScreenInput -Namespace RpaScreen | Out-Null; try { [RpaScreen.RpaScreenInput]::SetProcessDpiAwarenessContext([IntPtr](-4)) | Out-Null } catch {}";
 
 function screenCenter(selector: UiaSelector): { x: number; y: number; w: number; h: number } {
@@ -204,8 +400,8 @@ $deadline = [DateTime]::UtcNow.AddMilliseconds(${Math.max(0, Math.round(timeoutM
 $found = $null
 while ($null -eq $found) {
   try {
-    $parent = Find-RpaWindow ${psString(s.window ?? "")}
-    $found = Find-RpaElement $parent ${psString(s.name ?? "")} ${psString(s.id ?? "")} ${psString(s.type ?? "")} ${psString(s.className ?? "")} ${s.index}
+    ${elementResolveLines(s)}
+    $found = $el
   } catch {
     $err = $_.Exception.Message
     if ($err -like "window_not_found:*") { $found = $null }
@@ -239,8 +435,7 @@ function clickScript(t: UiaTarget): string {
 ${FIND_HELPER}
 $el = $null
 try {
-  $parent = Find-RpaWindow ${psString(s.window ?? "")}
-  $el = Find-RpaElement $parent ${psString(s.name ?? "")} ${psString(s.id ?? "")} ${psString(s.type ?? "")} ${psString(s.className ?? "")} ${s.index}
+  ${elementResolveLines(s)}
 } catch { $el = $null }
 if ($null -eq $el) { Write-Output "JSON_NOT_FOUND"; exit 0 }
 $rect = $el.Current.BoundingRectangle
@@ -335,10 +530,7 @@ export function setValueScript(t: UiaTarget, value: string): string {
   const s = t.selector;
   return `${UIA_USING}
 ${FIND_HELPER}
-try {
-  $parent = Find-RpaWindow ${psString(s.window ?? "")}
-  $el = Find-RpaElement $parent ${psString(s.name ?? "")} ${psString(s.id ?? "")} ${psString(s.type ?? "")} ${psString(s.className ?? "")} ${s.index}
-} catch { $el = $null }
+${elementResolveLines(s)}
 if ($null -eq $el) { Write-Output "JSON_NOT_FOUND"; exit 0 }
 $value = ${psString(value)}
 $method = "value_pattern"
@@ -372,10 +564,7 @@ function getValueScript(t: UiaTarget): string {
   const s = t.selector;
   return `${UIA_USING}
 ${FIND_HELPER}
-try {
-  $parent = Find-RpaWindow ${psString(s.window ?? "")}
-  $el = Find-RpaElement $parent ${psString(s.name ?? "")} ${psString(s.id ?? "")} ${psString(s.type ?? "")} ${psString(s.className ?? "")} ${s.index}
-} catch { $el = $null }
+${elementResolveLines(s)}
 if ($null -eq $el) { Write-Output "JSON_NOT_FOUND"; exit 0 }
 $value = $el.Current.Name
 try { $vp = $el.GetCurrentPattern([System.Windows.Automation.ValuePattern]::Pattern); $value = $vp.Value } catch {}
@@ -426,29 +615,37 @@ function requireFound(result: Record<string, unknown> | null, t: UiaTarget, acti
 
 const clickGround: CapabilityHandler = async (_ctx, node, args) => {
   const t = target(args, node);
+  if (t.selector.screen) return runScript(clickScreenScript(t), t);
   return runWithImplicitWait(t, "click", clickScript(t), runScript, args);
 };
 
 const setValueGround: CapabilityHandler = async (_ctx, node, args) => {
   const t = target(args, node);
   const value = String(args.value ?? args.text ?? "");
+  if (t.selector.screen) return runScript(setValueScreenScript(t, value), t);
   return runWithImplicitWait(t, "setValue", setValueScript(t, value), runScript, args);
 };
 
 const existsGround: CapabilityHandler = async (_ctx, node, args) => {
   const t = target(args, node);
+  if (t.selector.screen) return true;
   const found = await findElement(t, 1000, 150);
   return found !== null;
 };
 
 const getValueGround: CapabilityHandler = async (_ctx, node, args) => {
   const t = target(args, node);
+  if (t.selector.screen) {
+    const result = await runScript(getValueScreenScript(t), t);
+    return requireFound(result, t, "getValue").value ?? "";
+  }
   const result = await runWithImplicitWait(t, "getValue", getValueScript(t), runScript, args);
   return result.value ?? "";
 };
 
 const waitElementGround: CapabilityHandler = async (_ctx, node, args) => {
   const t = target(args, node);
+  if (t.selector.screen) return true;
   const timeout = Number(args.timeout ?? args.timeoutMs ?? 5000) || 5000;
   const gap = Number(args.time_gap ?? args.gapMs ?? 200) || 200;
   const found = await findElement(t, timeout, gap);
@@ -459,6 +656,39 @@ register("Element_ClickGround", clickGround);
 register("Element_SetValueGround", setValueGround);
 register("Element_ExistsGround", existsGround);
 register("Element_GetValueGround", getValueGround);
+const restoreWindowGround: CapabilityHandler = async (_ctx, node, args) => {
+  const t = target(args, node);
+  const lines = ["$hwnd = Restore-RpaWindow $el"];
+  if ([args.x, args.y, args.width, args.height].some((value) => value !== undefined && value !== null && String(value).trim() !== "")) {
+    const x = Math.round(Number(args.x ?? 60)) || 0;
+    const y = Math.round(Number(args.y ?? 60)) || 0;
+    const width = Math.round(Number(args.width ?? 0)) || 0;
+    const height = Math.round(Number(args.height ?? 0)) || 0;
+    lines.push("$hwnd = Move-RpaWindow $el " + x + " " + y + " " + width + " " + height);
+  }
+  return runWithImplicitWait(t, "restoreWindow", propertyValue([...lines, "Write-Output (@{ hwnd = [string]$hwnd; name = $el.Current.Name } | ConvertTo-Json -Compress)"], t), runScript, args);
+};
+
+const activateWindowGround: CapabilityHandler = async (_ctx, node, args) => {
+  const t = target(args, node);
+  return runWithImplicitWait(t, "activateWindow", propertyValue([
+    "$hwnd = Activate-RpaWindow $el",
+    "Write-Output (@{ hwnd = [string]$hwnd; name = $el.Current.Name } | ConvertTo-Json -Compress)"
+  ], t), runScript, args);
+};
+
+const moveWindowGround: CapabilityHandler = async (_ctx, node, args) => {
+  const t = target(args, node);
+  const x = Math.round(Number(args.x ?? 60)) || 0;
+  const y = Math.round(Number(args.y ?? 60)) || 0;
+  const width = Math.round(Number(args.width ?? 0)) || 0;
+  const height = Math.round(Number(args.height ?? 0)) || 0;
+  return runWithImplicitWait(t, "moveWindow", propertyValue([
+    "$hwnd = Move-RpaWindow $el " + x + " " + y + " " + width + " " + height,
+    "Write-Output (@{ hwnd = [string]$hwnd; name = $el.Current.Name } | ConvertTo-Json -Compress)"
+  ], t), runScript, args);
+};
+
 register("Element_WaitElementGround", waitElementGround);
 
 
@@ -486,8 +716,7 @@ function propertyValue(scriptLines: string[], t: UiaTarget): string {
 ${FIND_HELPER}
 $el = $null
 try {
-  $parent = Find-RpaWindow ${psString(s.window ?? "")}
-  $el = Find-RpaElement $parent ${psString(s.name ?? "")} ${psString(s.id ?? "")} ${psString(s.type ?? "")} ${psString(s.className ?? "")} ${s.index}
+  ${elementResolveLines(s)}
 } catch { $el = $null }
 if ($null -eq $el) { Write-Output "JSON_NOT_FOUND"; exit 0 }
 ${scriptLines.join("\n")}`;
@@ -709,3 +938,6 @@ register("Element_MouseToGround", mouseToGround);
 register("Element_ScrollIntoViewGround", scrollIntoViewGround);
 register("Element_ChildrenGround", childrenGroundHandler(false));
 register("Element_DesktopChildren", childrenGroundHandler(true));
+register("Element_RestoreWindowGround", restoreWindowGround);
+register("Element_ActivateWindowGround", activateWindowGround);
+register("Element_MoveWindowGround", moveWindowGround);
diff --git a/src/main/capabilities/index.ts b/src/main/capabilities/index.ts
index 14ca743..ff63b65 100644
--- a/src/main/capabilities/index.ts
+++ b/src/main/capabilities/index.ts
@@ -24,6 +24,7 @@ import "./word";
 import "./word-extended";
 import "./message";
 import "./vision";
+import "./cv-vision";
 import "./extended";
 import "./system-extended";
 import "./office-file-extended";
diff --git a/src/main/desktop-pick.ts b/src/main/desktop-pick.ts
index a5d9b37..4a5c516 100644
--- a/src/main/desktop-pick.ts
+++ b/src/main/desktop-pick.ts
@@ -3,9 +3,9 @@ import { promisify } from "node:util";
 import fs from "node:fs";
 
 import path from "node:path";
-import { app, BrowserWindow, screen } from "electron";
+import { app, BrowserWindow, desktopCapturer, globalShortcut, screen } from "electron";
 import { RuntimeError } from "./engine/types";
-import { chatCompletion, requireAiConfig } from "./capabilities/ai-service";
+import { detectUiElements, localUiModelPath } from "./local-vision/ui-detect-service";
 
 const run = promisify(execFile);
 
@@ -58,6 +58,8 @@ interface UiaPickOutput {
 }
 
 interface AiCandidate {
+  /** 候选稳定身份：覆盖层排序后仍按 ID 回填，禁止用排序后的数组下标当身份 */
+  candidateId: string;
   x: number;
   y: number;
   width: number;
@@ -92,12 +94,10 @@ function safeSelectorText(value: string | undefined): string {
   return String(value || "").replace(/[;\r\n]+/g, " ").trim();
 }
 
-export async function pickDesktopElement(options: { ai?: boolean } = {}): Promise<DesktopPickResult> {
-  if (options.ai) requireAiConfig({ vision: true });
+export async function pickDesktopElement(): Promise<DesktopPickResult> {
   const appWindows = BrowserWindow.getAllWindows().filter(w => !w.isDestroyed());
   for (const w of appWindows) w.minimize();
   try {
-    if (options.ai) return await runAiPick();
     const { stdout } = await run(
       "powershell.exe",
       ["-NoProfile", "-STA", "-ExecutionPolicy", "Bypass", "-File", pickerScriptPath()],
@@ -109,7 +109,9 @@ export async function pickDesktopElement(options: { ai?: boolean } = {}): Promis
       .map(value => Number(value))
       .every(value => Number.isFinite(value)) && Number(parsed.paneW) > 4 && Number(parsed.paneH) > 4;
     if (parsed.aiRequested && paneAiReady) return await runPaneAiPick(parsed);
-    if (parsed.aiRequested) return await runAiPick(parsed);
+    if (parsed.aiRequested) {
+      throw new RuntimeError("ai_pane_invalid", "AI 兜底识别仅支持包含有效区域的 Pane");
+    }
     const parts: string[] = [];
     if (parsed.window) parts.push("window=" + parsed.window);
     if (parsed.name) parts.push("name=" + parsed.name);
@@ -125,68 +127,121 @@ export async function pickDesktopElement(options: { ai?: boolean } = {}): Promis
 
 function paneAiOverlayHtml(): string {
   return `<!DOCTYPE html><html><head><meta charset='utf-8'><style>
-html,body{margin:0;width:100%;height:100%;overflow:hidden;background:transparent;cursor:default;font-family:'Microsoft YaHei',Arial,sans-serif}
-#frame{position:absolute;inset:0;display:none;border:1px solid rgba(56,189,248,.30);animation:framePulse 1.9s ease-in-out infinite;z-index:4}#frame i{position:absolute;width:30px;height:30px;border:3px solid #38bdf8;filter:drop-shadow(0 0 7px rgba(56,189,248,.85))}#frame i:nth-child(1){left:2px;top:2px;border-right:0;border-bottom:0;border-radius:10px 0 0 0}#frame i:nth-child(2){right:2px;top:2px;border-left:0;border-bottom:0;border-radius:0 10px 0 0}#frame i:nth-child(3){left:2px;bottom:2px;border-right:0;border-top:0;border-radius:0 0 0 10px}#frame i:nth-child(4){right:2px;bottom:2px;border-left:0;border-top:0;border-radius:0 0 10px 0}
-#scan{position:absolute;left:0;right:0;top:-20%;height:19%;display:none;background:linear-gradient(to bottom,rgba(56,189,248,0),rgba(56,189,248,.06) 48%,rgba(125,211,252,.20) 82%,rgba(224,242,254,.16));z-index:5;animation:scan 1.65s linear infinite;will-change:top}#scan b{position:absolute;left:3%;right:3%;bottom:-2px;height:3px;border-radius:999px;background:linear-gradient(90deg,rgba(56,189,248,0),#7dd3fc 14%,#fff 50%,#7dd3fc 86%,rgba(56,189,248,0));box-shadow:0 0 10px #38bdf8,0 0 28px rgba(56,189,248,.78),0 0 54px rgba(56,189,248,.38)}#scan b:after{content:'';position:absolute;left:0;right:0;top:2px;height:9px;background:linear-gradient(to bottom,rgba(56,189,248,.18),rgba(56,189,248,0))}
-#tip{position:absolute;left:50%;top:18px;transform:translateX(-50%);display:none;padding:8px 15px;border-radius:999px;color:#f8fafc;background:rgba(2,20,40,.78);border:1px solid rgba(125,211,252,.42);box-shadow:0 8px 28px rgba(0,0,0,.28),inset 0 1px 0 rgba(255,255,255,.09);font-size:13px;line-height:1;white-space:nowrap;z-index:2147483647}#tip:before{content:'';display:inline-block;width:7px;height:7px;margin-right:8px;border-radius:50%;background:#38bdf8;box-shadow:0 0 8px #38bdf8;vertical-align:1px}body.scanning #tip:before{animation:tipPulse 1.1s ease-in-out infinite}
-.candidate{position:absolute;border:1px solid rgba(255,255,255,.28);background:transparent;cursor:pointer;box-sizing:border-box}
-.candidate:hover,.candidate.active{border:3px solid #ff2d2d;box-shadow:0 0 0 1px rgba(255,255,255,.45);z-index:2147483646}
-.badge{position:absolute;display:none;padding:3px 7px;border-radius:5px;color:#fff;background:rgba(190,18,18,.95);border:1px solid rgba(255,220,220,.82);font-size:12px;line-height:16px;white-space:nowrap;pointer-events:none;box-shadow:0 4px 12px rgba(0,0,0,.28);z-index:2147483647}.badge.active{display:block}
-@keyframes scan{from{top:-20%}to{top:100%}}@keyframes framePulse{0%,100%{border-color:rgba(56,189,248,.24)}50%{border-color:rgba(56,189,248,.58)}}@keyframes tipPulse{0%,100%{opacity:.45}50%{opacity:1}}
+html,body{margin:0;width:100%;height:100%;overflow:hidden;background:transparent;font-family:'Microsoft YaHei',Arial,sans-serif}
+#veil{position:absolute;inset:0;background:rgba(2,10,28,.035);opacity:0;transition:opacity .18s ease;z-index:3;pointer-events:none}
+body.scanning #veil{opacity:1}
+#paneFrame{position:absolute;inset:0;border:2px solid #ff3b30;box-shadow:0 0 0 1px rgba(255,255,255,.34),0 0 22px rgba(255,59,48,.24);z-index:4;pointer-events:none}
+#aiFrame{position:absolute;inset:6px;border:1px solid rgba(56,189,248,.26);border-radius:12px;animation:aiPulse 2s ease-in-out infinite;z-index:5;pointer-events:none}
+#aiFrame i{position:absolute;width:22px;height:22px;border:2px solid #38bdf8;filter:drop-shadow(0 0 6px rgba(56,189,248,.72))}
+#aiFrame i:nth-child(1){left:-1px;top:-1px;border-right:0;border-bottom:0;border-radius:10px 0 0 0}
+#aiFrame i:nth-child(2){right:-1px;top:-1px;border-left:0;border-bottom:0;border-radius:0 10px 0 0}
+#aiFrame i:nth-child(3){left:-1px;bottom:-1px;border-right:0;border-top:0;border-radius:0 0 0 10px}
+#aiFrame i:nth-child(4){right:-1px;bottom:-1px;border-left:0;border-top:0;border-radius:0 0 10px 0}
+#scan{position:absolute;left:0;right:0;top:0;height:1px;display:none;background:linear-gradient(90deg,rgba(56,189,248,0),rgba(125,211,252,.72) 16%,rgba(255,255,255,.94) 50%,rgba(125,211,252,.72) 84%,rgba(56,189,248,0));box-shadow:0 0 10px rgba(56,189,248,.78);animation:scan 1.35s cubic-bezier(.4,0,.6,1) infinite;z-index:6;pointer-events:none}
+#scan:before{content:'';position:absolute;left:0;right:0;top:1px;height:58px;background:linear-gradient(to bottom,rgba(56,189,248,.13),rgba(56,189,248,0))}
+#tip{position:absolute;left:50%;top:14px;transform:translateX(-50%);display:none;min-height:30px;padding:0 13px;border-radius:16px;color:#f8fafc;background:rgba(4,18,34,.84);border:1px solid rgba(125,211,252,.34);box-shadow:0 8px 24px rgba(0,0,0,.24),inset 0 1px 0 rgba(255,255,255,.06);font-size:12.5px;line-height:28px;white-space:nowrap;z-index:2147483647;pointer-events:none}
+#tip:before{content:'';display:inline-block;width:6px;height:6px;margin-right:8px;border-radius:50%;background:#38bdf8;box-shadow:0 0 8px #38bdf8;vertical-align:2px}
+body.scanning #tip:before{animation:tipPulse 1.15s ease-in-out infinite}
+body.debug .candidate{border:1px solid rgba(255,59,48,.74)}
+.candidate{position:absolute;background:transparent;cursor:pointer;box-sizing:border-box;z-index:100}
+.candidate{border:3px solid transparent}.candidate.active{border-color:#ff3b30;box-shadow:0 0 0 1px rgba(255,255,255,.38);z-index:101}
+.badge{position:absolute;display:none;padding:3px 7px;border-radius:5px;color:#fff;background:rgba(190,18,18,.94);border:1px solid rgba(255,220,220,.78);font-size:12px;line-height:16px;white-space:nowrap;pointer-events:none;box-shadow:0 4px 12px rgba(0,0,0,.24);z-index:2147483647}
+.badge.active{display:block;z-index:102}
+@keyframes scan{from{top:0}to{top:100%}}
+@keyframes aiPulse{0%,100%{border-color:rgba(56,189,248,.18)}50%{border-color:rgba(56,189,248,.46)}}
+@keyframes tipPulse{0%,100%{opacity:.4}50%{opacity:1}}
 </style></head><body>
-<div id=frame><i></i><i></i><i></i><i></i></div><div id=scan><b></b></div><div id=tip>AI 识别中…</div>
+<div id=veil></div><div id=paneFrame></div><div id=aiFrame><i></i><i></i><i></i><i></i></div><div id=scan></div><div id=tip>AI 识别中…</div>
 <script>
-var paneW=1,paneH=1;
-function startScan(label){paneW=window.__paneW||document.documentElement.clientWidth;paneH=window.__paneH||document.documentElement.clientHeight;document.getElementById('frame').style.display='block';document.getElementById('scan').style.display='block';document.getElementById('tip').style.display='block';document.getElementById('tip').textContent=label||'AI 识别中…';document.body.classList.add('scanning')}
+var paneW=1,paneH=1,lastMouse=null,lastCid=null;
+function showStatus(label){document.getElementById('tip').style.display='block';document.getElementById('tip').textContent=label||'AI 识别中…';}
+function startScan(label){
+  paneW=window.__paneW||document.documentElement.clientWidth;paneH=window.__paneH||document.documentElement.clientHeight;
+  document.body.classList.add('scanning');
+  document.getElementById('scan').style.display='block';
+  document.getElementById('aiFrame').style.display='block';
+  showStatus(label||'AI 识别中…');
+}
+function clearActive(){
+  if(!lastCid)return;
+  Array.prototype.forEach.call(document.querySelectorAll('.candidate.active,.badge.active'),function(n){n.classList.remove('active')});
+  lastCid=null;window.__hoverCid=null;window.__hoverCandidate=null;
+  if(window.__candidatesReady)logHoverState('clear');
+}
+function activateBox(box){
+  var cid=box.dataset.cid;
+  if(lastCid===cid)return;
+  clearActive();
+  var index=Number(box.dataset.index),label=document.getElementById('badge-'+index);
+  box.classList.add('active');if(label)label.classList.add('active');
+  lastCid=cid;window.__hoverCid=cid;window.__hoverCandidate=window.__candidateMap[cid];
+  logHoverState('active');
+}
+function syncHover(){
+  if(!window.__candidatesReady)return;
+  var event=lastMouse;if(!event){clearActive();return}
+  // 当前框未离开前保持选中，避免重叠候选因 1px 抖动来回抢高亮。
+  var current=lastCid?document.querySelector('.candidate[data-cid="'+lastCid+'"]'):null;
+  if(current){
+    var r=current.getBoundingClientRect();
+    if(event.clientX>=r.left-2&&event.clientX<=r.right+2&&event.clientY>=r.top-2&&event.clientY<=r.bottom+2)return;
+  }
+  var hit=null,hitArea=Infinity;
+  Array.prototype.forEach.call(document.querySelectorAll('.candidate'),function(box){
+    var rect=box.getBoundingClientRect();
+    if(event.clientX<rect.left||event.clientX>rect.right||event.clientY<rect.top||event.clientY>rect.bottom)return;
+    var area=rect.width*rect.height;
+    if(area<hitArea){hitArea=area;hit=box}
+  });
+  if(hit)activateBox(hit);else clearActive();
+}
+function logHoverState(state){
+  console.log('pane-ai hover ' + state + ' mouse=' + (lastMouse?lastMouse.clientX+','+lastMouse.clientY:'none') +
+    ' cid=' + (lastCid||'none') + ' candidates=' + document.querySelectorAll('.candidate').length);
+}
+window.__toggleDebug=function(){
+  document.body.classList.toggle('debug');
+  var debug=document.body.classList.contains('debug');
+  showStatus(debug?'开发模式：F8 关闭候选框':(window.__candidatesReady?'AI 完成：移动高亮；单击选中，再次单击或空格回填；Esc 取消':'开发模式已开启，识别完成后显示全部候选框'));
+};
 function showCandidates(items){
   paneW=window.__paneW||document.documentElement.clientWidth;paneH=window.__paneH||document.documentElement.clientHeight;
-  document.getElementById('scan').style.display='none';document.getElementById('frame').style.display='none';document.getElementById('tip').style.display='block';
-  document.getElementById('tip').textContent='AI 识别完成：移动鼠标高亮；单击选中，再次单击或空格回填；Esc 取消';
+  document.getElementById('scan').style.display='none';
+  document.getElementById('aiFrame').style.display='none';
   document.body.classList.remove('scanning');document.body.style.cursor='pointer';
-  window.__hover=null;window.__selected=null;window.__confirmed=null;window.__readyAt=Date.now()+600;
-  items.sort(function(a,b){return a.width*a.height-b.width*b.height}).forEach(function(c,index){
-    var box=document.createElement('div');box.className='candidate';box.__index=index;
+  window.__candidateMap={};window.__candidatesReady=false;clearActive();
+  items.sort(function(a,b){return a.y-b.y||a.x-b.x}).forEach(function(c,index){
+    window.__candidateMap[c.candidateId]=c;
+    var box=document.createElement('div');box.className='candidate';box.dataset.cid=c.candidateId;box.dataset.index=index;
     box.style.left=(c.x/paneW*100)+'%';box.style.top=(c.y/paneH*100)+'%';
     box.style.width=(c.width/paneW*100)+'%';box.style.height=(c.height/paneH*100)+'%';
-    function activate(){
-      Array.prototype.forEach.call(document.querySelectorAll('.candidate,.badge'),function(n){n.classList.remove('active')});
-      box.classList.add('active');var label=document.getElementById('badge-'+index);if(label)label.classList.add('active');
-      window.__hover=index;window.__hoverCandidate=c;
-    }
-    window.confirmCandidate=function(index,c){
-      var tip=document.getElementById('tip');
-      tip.textContent='已选中：'+(c.name||'控件')+'；再次单击或空格回填，Esc 取消';
-      window.__confirmed=index;window.__selected=index;
-    };
-    function deactivate(){
-      box.classList.remove('active');var label=document.getElementById('badge-'+index);if(label)label.classList.remove('active');
-      if(window.__hover===index){window.__hover=null;window.__hoverCandidate=null}
-    }
-    box.addEventListener('mouseenter',activate);box.addEventListener('mousemove',activate);box.addEventListener('mouseleave',deactivate);
-    box.addEventListener('mousedown',function(e){
-      e.preventDefault();e.stopPropagation();activate();
-      if(Date.now()<(window.__readyAt||0)){console.log('pane-ai ignore early click',index);return}
-      if(window.__confirmed!==index){console.log('pane-ai first select',index);window.confirmCandidate(index,c);return}
-      console.log('pane-ai confirm click',index);window.__selected=index;window.__done=true;
+    var label=document.createElement('div');label.className='badge';label.id='badge-'+index;label.textContent='AI识别 · '+(c.controlType||'控件');
+    label.style.left=(c.x/paneW*100)+'%';label.style.top=(Math.max(0,c.y-24)/paneH*100)+'%';
+    box.addEventListener('click',function(event){
+      event.stopPropagation();if(Date.now()<window.__readyAt)return;
+      activateBox(box);
+      if(window.__selectedCid!==c.candidateId){window.__selectedCid=c.candidateId;showStatus('已选中：'+(c.controlType||'控件')+'；再次单击或空格回填');return}
+      window.__confirmedCid=c.candidateId;window.__done=true;
     });
-    document.body.appendChild(box);
-    var yTop=c.y/paneH*100,yBottom=(c.y+c.height)/paneH*100,leftPct=Math.max(0,Math.min(72,c.x/paneW*100));
-    var badge=document.createElement('div');badge.id='badge-'+index;badge.className='badge';
-    badge.textContent='AI识别 · '+(c.name||'控件');
-    badge.style.left=leftPct+'%';
-    badge.style.top=(yTop<9?'calc('+yBottom+'% + 8px)':'calc('+yTop+'% - 32px)');
-    document.body.appendChild(badge);
+    document.body.appendChild(box);document.body.appendChild(label);
   });
+  window.__candidatesReady=true;window.__readyAt=Date.now()+450;
+  showStatus(document.body.classList.contains('debug')?'开发模式：F8 关闭候选框':'AI 完成：移动高亮；单击选中，再次单击或空格回填；Esc 取消');
+  syncHover();
 }
-document.addEventListener('keydown',function(e){
-  if(e.key==='Escape'){window.__cancelled=true}
-  if(e.key===' '&&window.__hoverCandidate&&!window.__done){
-    e.preventDefault();
-    if(Date.now()<(window.__readyAt||0)){console.log('pane-ai ignore early space');return}
-    if(window.__confirmed!==window.__hover){window.confirmCandidate(window.__hover,window.__hoverCandidate);return}
-    window.__selected=window.__hover;window.__done=true;
+document.addEventListener('pointermove',function(event){lastMouse=event;syncHover()});
+document.addEventListener('keydown',function(event){
+  if(event.key==='F8'||event.code==='F8'){event.preventDefault();window.__toggleDebug();return}
+  if(event.key==='Escape'){window.__cancelled=true;return}
+  if(event.key===' '||event.code==='Space'){event.preventDefault();
+    if(window.__confirmedCid!==window.__hoverCid){window.__confirmedCid=window.__hoverCid;window.__done=true}
   }
 });
+document.addEventListener('mousedown',function(event){
+  var candidate=event.target&&event.target.classList&&event.target.classList.contains('candidate');
+  if(candidate||Date.now()<(window.__readyAt||0))return;
+  if(window.__selectedCid&&window.__hoverCid===window.__selectedCid){window.__confirmedCid=window.__hoverCid;window.__done=true}
+});
 </script></body></html>`;
 }
 
@@ -236,160 +291,168 @@ async function createPaneAiOverlay(rect: { x: number; y: number; w: number; h: n
   return win;
 }
 
-async function capturePaneForAi(rect: { x: number; y: number; w: number; h: number }): Promise<AiCaptureOutput> {
-  const outDir = path.join(app.getPath("userData"), "ai-picks");
-  fs.mkdirSync(outDir, { recursive: true });
-  const imagePath = path.join(outDir, "ai-pane-" + Date.now() + ".png");
-  const { stdout } = await run(
-    "powershell.exe",
-    [
-      "-NoProfile", "-STA", "-ExecutionPolicy", "Bypass", "-File", pickerScriptPath(),
-      "-Mode", "capture", "-CaptureExact",
-      "-CaptureX", String(rect.x), "-CaptureY", String(rect.y),
-      "-CaptureWidth", String(rect.w), "-CaptureHeight", String(rect.h),
-      "-CapturePath", imagePath
-    ],
-    { timeout: 30000, windowsHide: true, maxBuffer: 2 * 1024 * 1024 }
-  );
-  const parsed = jsonLine<AiCaptureOutput>(stdout, "{");
-  if (!parsed.imagePath || !fs.existsSync(parsed.imagePath)) {
-    throw new RuntimeError("ai_capture_failed", "AI 拾取截图失败");
-  }
-  return parsed;
-}
+async function capturePaneForAi(rect: { x: number; y: number; w: number; h: number }, overlay: BrowserWindow): Promise<AiCaptureOutput> {
+  const centerDip = screen.screenToDipPoint({ x: rect.x + rect.w / 2, y: rect.y + rect.h / 2 });
+  const display = displayForPoint(centerDip.x, centerDip.y);
+  const scale = display.scaleFactor || 1;
+  const displayPhysicalX = Math.round(display.bounds.x * scale);
+  const displayPhysicalY = Math.round(display.bounds.y * scale);
+  // UIA/Desktop Picker 返回的是物理像素坐标；截图裁剪也必须直接使用物理像素。
+  // 这里不能把 cropX/cropY 再乘 actualScale，否则会重复缩放并导致候选框整体偏移。
+  const cropX = rect.x - displayPhysicalX;
+  const cropY = rect.y - displayPhysicalY;
+  const thumbnailWidth = Math.round(display.bounds.width * scale);
+  const thumbnailHeight = Math.round(display.bounds.height * scale);
 
-function parseAiCandidates(text: string, cropW: number, cropH: number): AiCandidate[] {
-  const fence = String.fromCharCode(96);
-  const fencePattern = new RegExp(fence + "{3}(?:json)?" + fence + "{3}", "gi");
-  const cleaned = text.replace(fencePattern, "").trim();
-  const start = cleaned.indexOf("{");
-  const arrayStart = cleaned.indexOf("[");
-  const objectStart = start >= 0 && (start < arrayStart || arrayStart < 0) ? start : arrayStart;
-  const end = Math.max(cleaned.lastIndexOf("}"), cleaned.lastIndexOf("]"));
-  if (objectStart < 0 || end <= objectStart) throw new RuntimeError("ai_response_invalid", "AI 视觉模型没有返回 JSON");
-  const raw = JSON.parse(cleaned.slice(objectStart, end + 1)) as Record<string, unknown>;
-  const list = Array.isArray(raw) ? raw : Array.isArray(raw.candidates) ? raw.candidates : [];
-  const result: AiCandidate[] = [];
-  for (const item of list.slice(0, 40)) {
-    const record = (item || {}) as Record<string, unknown>;
-    const num = (...keys: string[]): number => {
-      for (const key of keys) {
-        const value = Number(record[key]);
-        if (Number.isFinite(value)) return value;
-      }
-      return NaN;
-    };
-    let x = num("x", "left", "x1");
-    let y = num("y", "top", "y1");
-    let width = num("width", "w");
-    let height = num("height", "h");
-    const x2 = num("x2", "right");
-    const y2 = num("y2", "bottom");
-    if (!Number.isFinite(width) && Number.isFinite(x2)) width = x2 - x;
-    if (!Number.isFinite(height) && Number.isFinite(y2)) height = y2 - y;
-    if (![x, y, width, height].every(Number.isFinite) || width < 6 || height < 6) continue;
-    x = Math.max(0, Math.min(cropW - width, x));
-    y = Math.max(0, Math.min(cropH - height, y));
-    width = Math.min(width, cropW - x);
-    height = Math.min(height, cropH - y);
-    if (width * height >= cropW * cropH * 0.97) continue;
-    const confidence = Number(record.confidence);
-    result.push({
-      x: Math.round(x), y: Math.round(y), width: Math.round(width), height: Math.round(height),
-      name: String(record.name ?? "AI识别控件").slice(0, 80) || "AI识别控件",
-      controlType: String(record.controlType ?? "Control").slice(0, 40) || "Control",
-      confidence: Number.isFinite(confidence) ? Math.max(0, Math.min(1, confidence)) : 0.72
+  // 只为避开拾取窗口本身；推理和动画继续在后台进行。
+  overlay.hide();
+  await new Promise(resolve => setTimeout(resolve, 60));
+  try {
+    const sources = await desktopCapturer.getSources({
+      types: ["screen"],
+      thumbnailSize: { width: thumbnailWidth, height: thumbnailHeight }
     });
+    overlay.show();
+    const source = sources.find(item => item.display_id === String(display.id)) || sources[0];
+    if (!source || source.thumbnail.isEmpty()) {
+      throw new RuntimeError("ai_capture_failed", "AI 拾取截图失败");
+    }
+
+    const thumbnail = source.thumbnail;
+    const actualScale = thumbnail.getSize().width / Math.max(1, display.bounds.width);
+    const x = Math.max(0, Math.min(thumbnail.getSize().width - 1, cropX));
+    const y = Math.max(0, Math.min(thumbnail.getSize().height - 1, cropY));
+    const width = Math.max(1, Math.min(thumbnail.getSize().width - x, Math.round(rect.w * actualScale)));
+    const height = Math.max(1, Math.min(thumbnail.getSize().height - y, Math.round(rect.h * actualScale)));
+    const cropped = thumbnail.crop({ x, y, width, height });
+    const image = cropped.toPNG();
+    const outDir = path.join(app.getPath("userData"), "ai-picks");
+    fs.mkdirSync(outDir, { recursive: true });
+    const imagePath = path.join(outDir, `ai-pane-${Date.now()}.png`);
+    await fs.promises.writeFile(imagePath, image);
+    return {
+      cursorX: rect.x,
+      cursorY: rect.y,
+      cropX: rect.x,
+      cropY: rect.y,
+      cropW: width,
+      cropH: height,
+      markerX: 0,
+      markerY: 0,
+      imagePath
+    };
+  } catch (error) {
+    overlay.show();
+    throw error;
   }
-  if (!result.length) throw new RuntimeError("ai_response_invalid", "AI 没有识别到可用候选区域");
-  return result;
 }
 
 async function runPaneAiPick(context: UiaPickOutput): Promise<DesktopPickResult> {
-  requireAiConfig({ vision: true });
+  // 本地小模型是 Pane AI 的主链路；这里不要求用户配置远程 VLM。
+  const localModelPath = localUiModelPath();
+  console.log("[pane-ai] local model", localModelPath);
   const pane = paneAiRect(context);
   const win = await createPaneAiOverlay(pane);
+  const startedAt = Date.now();
+  // F8 用全局快捷键，避免键盘焦点停在目标应用时 overlay 收不到按键。
+  const debugHotkeyRegistered = globalShortcut.register("F8", () => {
+    void win.webContents.executeJavaScript("window.__toggleDebug && window.__toggleDebug()", true).catch(() => {});
+  });
 
   try {
-    // Let the visible scan animation complete over the actual Pane before taking the screenshot.
-    await new Promise(resolve => setTimeout(resolve, 950));
-    win.hide();
-    await new Promise(resolve => setTimeout(resolve, 240));
-    const capture = await capturePaneForAi(pane);
-    win.show();
-    win.focus();
-    await win.webContents.executeJavaScript("startScan('AI 理解中…');");
+    // 稍等 overlay 完成首轮绘制，但不再等一整套“表演动画”。
+    await new Promise(resolve => setTimeout(resolve, 350));
+    const detection = capturePaneForAi(pane, win)
+      .then(async capture => {
+        const image = await fs.promises.readFile(capture.imagePath);
+        const inferenceStarted = Date.now();
+        const boxes = await detectUiElements({ image, width: capture.cropW, height: capture.cropH });
+        console.log("[pane-ai] local inference", boxes.length, Date.now() - inferenceStarted + "ms");
+        return { capture, boxes };
+      });
 
-    const image = await fs.promises.readFile(capture.imagePath);
-    let aiSettled = false;
-    const aiRequest = chatCompletion([
-      {
-        role: "system",
-        content: "你是桌面 RPA 视觉控件分析器。只输出 JSON，不输出解释或 Markdown。"
-      },
-      {
-        role: "user",
-        content: [
-          {
-            type: "text",
-            text: [
-              "这是目标 Pane 区域的截图，尺寸 " + capture.cropW + "x" + capture.cropH + " 像素。",
-              "请把这个 Pane 里的可用控件/内容块拆分成多个候选矩形，最多返回 30 个，按从上到下、从左到右排序。",
-              "优先识别按钮、输入框、列表行、卡片、菜单项、链接和明显可操作区域；不要把整个 Pane 当作候选。",
-              "只返回 JSON：{\"candidates\":[{\"x\":number,\"y\":number,\"width\":number,\"height\":number,\"name\":string,\"controlType\":string,\"confidence\":number}]}",
-              "坐标全部是相对这张截图的整数像素坐标；x/y 是左上角，width/height 也是像素。"
-            ].join("\n")
-          },
-          { type: "image_url", image_url: { url: "data:image/png;base64," + image.toString("base64") } }
-        ]
+    let settled = false;
+    const tickStatus = async () => {
+      const elapsed = Date.now() - startedAt;
+      const label = elapsed < 2600
+        ? "AI 识别中…"
+        : elapsed < 6000
+          ? "AI 理解中…"
+          : "正在整理候选区域…";
+      try {
+        await win.webContents.executeJavaScript("showStatus(" + JSON.stringify(label) + ");", true);
+      } catch {
+        // Overlay 销毁时状态更新可以安全忽略。
       }
-    ], { vision: true, temperature: 0 }).finally(() => { aiSettled = true; });
+    };
+    const statusTimer = setInterval(() => void tickStatus(), 1500);
 
     const cancelWatcher = (async () => {
-      while (!aiSettled) {
+      while (!settled) {
         try {
           const state = await win.webContents.executeJavaScript("JSON.stringify({cancelled:window.__cancelled===true})", true);
           if (JSON.parse(state).cancelled) throw new RuntimeError("pick_cancelled", "已取消 AI 拾取");
         } catch (error) {
           if (error instanceof RuntimeError && error.code === "pick_cancelled") throw error;
         }
-        await new Promise(resolve => setTimeout(resolve, 100));
+        await new Promise(resolve => setTimeout(resolve, 80));
       }
     })();
 
-    let aiText: string;
+    let detected: { capture: AiCaptureOutput; boxes: Awaited<ReturnType<typeof detectUiElements>> };
     try {
-      aiText = await Promise.race([aiRequest, cancelWatcher]) as string;
+      detected = await Promise.race([detection, cancelWatcher]) as { capture: AiCaptureOutput; boxes: Awaited<ReturnType<typeof detectUiElements>> };
     } finally {
-      aiSettled = true;
+      settled = true;
+      clearInterval(statusTimer);
+    }
+
+    const candidates: AiCandidate[] = detected.boxes.map((box, index) => ({
+      candidateId: "local-" + index,
+      x: box.x,
+      y: box.y,
+      width: box.width,
+      height: box.height,
+      name: box.category,
+      controlType: box.category,
+      confidence: box.confidence
+    }));
+    if (!candidates.length) {
+      throw new RuntimeError("ai_response_invalid", "本地 UI 模型没有识别到可用候选区域");
     }
 
-    const candidates = parseAiCandidates(aiText, capture.cropW, capture.cropH);
     const payload = JSON.stringify(candidates);
-    console.log("[pane-ai] candidates", candidates.length, candidates.map(item => ({ name: item.name, x: item.x, y: item.y, width: item.width, height: item.height })));
-    await win.webContents.executeJavaScript("window.__done=false;window.__cancelled=false;window.__selected=null;window.__hover=null;window.__hoverCandidate=null;window.__confirmed=null;window.__readyAt=Date.now()+600;showCandidates(" + payload + ");");
+    await win.show();
+    win.focus();
+    await win.webContents.executeJavaScript(
+      "window.__done=false;window.__cancelled=false;window.__selectedCid=null;window.__hoverCid=null;window.__hoverCandidate=null;window.__confirmedCid=null;window.__readyAt=Date.now()+500;showCandidates(" + payload + ");",
+      true
+    );
 
     const deadline = Date.now() + 180000;
-    let selectedIndex: number | null = null;
+    let selectedId: string | null = null;
     while (Date.now() < deadline) {
       await new Promise(resolve => setTimeout(resolve, 80));
       const state = await win.webContents.executeJavaScript(
-        "JSON.stringify({done:window.__done===true,cancelled:window.__cancelled===true,selected:window.__selected})", true
+        "JSON.stringify({done:window.__done===true,cancelled:window.__cancelled===true,selectedCid:window.__selectedCid||null})",
+        true
       );
-      const parsed = JSON.parse(state) as { done: boolean; cancelled: boolean; selected: number | null };
+      const parsed = JSON.parse(state) as { done: boolean; cancelled: boolean; selectedCid: string | null };
       if (parsed.cancelled) throw new RuntimeError("pick_cancelled", "已取消 AI 拾取");
-      if (parsed.done && Number.isInteger(parsed.selected)) {
-        const confirmedIndex = Number(parsed.selected);
-        console.log("[pane-ai] confirmed", confirmedIndex, candidates[confirmedIndex]);
-        selectedIndex = confirmedIndex;
-        break;
+      if (parsed.done && parsed.selectedCid) {
+        const box = candidates.find(item => item.candidateId === parsed.selectedCid);
+        console.log("[pane-ai] confirmed", parsed.selectedCid, box);
+        if (box) {
+          selectedId = box.candidateId;
+          break;
+        }
       }
     }
-    if (selectedIndex === null || !candidates[selectedIndex]) throw new RuntimeError("pick_cancelled", "AI 拾取超时");
+    if (!selectedId) throw new RuntimeError("pick_cancelled", "AI 拾取超时");
 
-    const box = candidates[selectedIndex];
-    const x = capture.cropX + box.x;
-    const y = capture.cropY + box.y;
+    const box = candidates.find(item => item.candidateId === selectedId)!;
+    const x = detected.capture.cropX + box.x;
+    const y = detected.capture.cropY + box.y;
     const selectorParts = [
       "screen=" + [x, y, box.width, box.height].map(Math.round).join(","),
       context.window ? "window=" + safeSelectorText(context.window) : "",
@@ -404,37 +467,15 @@ async function runPaneAiPick(context: UiaPickOutput): Promise<DesktopPickResult>
       controlType: box.controlType,
       x, y, w: box.width, h: box.height,
       confidence: box.confidence,
-      screenshotPath: capture.imagePath,
+      screenshotPath: detected.capture.imagePath,
       selector: selectorParts.join(";")
     };
   } finally {
+    if (debugHotkeyRegistered) globalShortcut.unregister("F8");
     if (!win.isDestroyed()) win.destroy();
   }
 }
 
-function aiOverlayHtml(): string {
-  return `<!DOCTYPE html><html><head><meta charset='utf-8'><style>
-html,body{margin:0;width:100%;height:100%;overflow:hidden;background:rgba(3,20,45,.12);cursor:crosshair;font-family:'Microsoft YaHei',Arial,sans-serif}
-#tip{position:fixed;top:28px;left:50%;transform:translateX(-50%);padding:10px 20px;border-radius:8px;color:#f8fafc;background:rgba(2,14,34,.82);border:1px solid rgba(96,165,250,.55);font-size:15px;box-shadow:0 8px 30px rgba(0,0,0,.25)}
-#scan{position:absolute;left:0;right:0;top:-240px;height:240px;display:none;background:linear-gradient(to bottom,rgba(59,130,246,0),rgba(59,130,246,.12),rgba(147,197,253,.46),rgba(59,130,246,.10),rgba(59,130,246,0));box-shadow:0 0 38px rgba(96,165,250,.88);animation:scan 1.7s linear infinite}
-#point{position:absolute;width:46px;height:46px;border:2px solid #38bdf8;border-radius:8px;display:none;box-shadow:0 0 20px rgba(56,189,248,.9)}
-body.scanning{background:rgba(3,20,45,.03)}
-body.scanning #scan{display:block}
-body.scanning #tip{font-size:13px}
-@keyframes scan{from{top:-240px}to{top:100%}}
-</style></head><body><div id=scan></div><div id=point></div><div id=tip>AI 拾取：单击要识别的控件；Esc 取消</div><script>
-document.addEventListener('mousedown', function(e) {
-  if (document.body.classList.contains('scanning')) return;
-  var p=document.getElementById('point');
-  p.style.left=(e.clientX-23)+'px';p.style.top=(e.clientY-23)+'px';p.style.display='block';
-  document.getElementById('tip').textContent='AI 视觉识别中…自上而下扫描画面';
-  document.body.classList.add('scanning');
-  window.__point={x:e.clientX,y:e.clientY};window.__done=true;
-});
-document.addEventListener('keydown',function(e){if(e.key==='Escape'){window.__cancelled=true;}});
-</script></body></html>`;
-}
-
 function displayForPoint(x: number, y: number) {
   return screen.getAllDisplays().find(display => {
     const b = display.bounds;
@@ -442,216 +483,6 @@ function displayForPoint(x: number, y: number) {
   }) || screen.getPrimaryDisplay();
 }
 
-async function createAiOverlay(): Promise<{ win: BrowserWindow; display: ReturnType<typeof screen.getPrimaryDisplay> }> {
-  const cursor = screen.getCursorScreenPoint();
-  const display = displayForPoint(cursor.x, cursor.y);
-  const win = new BrowserWindow({
-    x: display.bounds.x,
-    y: display.bounds.y,
-    width: display.bounds.width,
-    height: display.bounds.height,
-    frame: false,
-    transparent: true,
-    resizable: false,
-    movable: false,
-    minimizable: false,
-    maximizable: false,
-    alwaysOnTop: true,
-    skipTaskbar: true,
-    show: false,
-    webPreferences: { contextIsolation: true, nodeIntegration: false, sandbox: true }
-  });
-  win.setAlwaysOnTop(true, "screen-saver");
-  await win.loadURL("data:text/html;charset=utf-8," + encodeURIComponent(aiOverlayHtml()));
-  win.show();
-  win.focus();
-  return { win, display };
-}
-
-async function captureAiAtPoint(physicalX: number, physicalY: number): Promise<AiCaptureOutput> {
-  const outDir = path.join(app.getPath("userData"), "ai-picks");
-  fs.mkdirSync(outDir, { recursive: true });
-  const imagePath = path.join(outDir, "ai-pick-" + Date.now() + ".png");
-  const { stdout } = await run(
-    "powershell.exe",
-    [
-      "-NoProfile", "-STA", "-ExecutionPolicy", "Bypass", "-File", pickerScriptPath(),
-      "-Mode", "capture", "-CaptureX", String(physicalX), "-CaptureY", String(physicalY),
-      "-CaptureWidth", "1000", "-CaptureHeight", "700", "-CapturePath", imagePath
-    ],
-    { timeout: 30000, windowsHide: true, maxBuffer: 2 * 1024 * 1024 }
-  );
-  const parsed = jsonLine<AiCaptureOutput>(stdout, "{");
-  if (!parsed.imagePath || !fs.existsSync(parsed.imagePath)) {
-    throw new RuntimeError("ai_capture_failed", "AI 拾取截图失败");
-  }
-  return parsed;
-}
-
-function parseAiBox(text: string, cropW: number, cropH: number) {
-  const fence = String.fromCharCode(96);
-  const fencePattern = new RegExp(fence + "{3}(?:json)?" + fence + "{3}", "gi");
-  const cleaned = text.replace(fencePattern, "").trim();
-  const start = cleaned.indexOf("{");
-  const end = cleaned.lastIndexOf("}");
-  if (start < 0 || end <= start) throw new RuntimeError("ai_response_invalid", "AI 视觉模型没有返回 JSON");
-  const raw = JSON.parse(cleaned.slice(start, end + 1)) as Record<string, unknown>;
-  const num = (...keys: string[]): number => {
-    for (const key of keys) {
-      const value = Number(raw[key]);
-      if (Number.isFinite(value)) return value;
-    }
-    return NaN;
-  };
-  const x1 = num("x", "left", "x1");
-  const y1 = num("y", "top", "y1");
-  let width = num("width", "w");
-  let height = num("height", "h");
-  const x2 = num("x2", "right");
-  const y2 = num("y2", "bottom");
-  if (!Number.isFinite(width) && Number.isFinite(x2)) width = x2 - x1;
-  if (!Number.isFinite(height) && Number.isFinite(y2)) height = y2 - y1;
-  if (![x1, y1].every(Number.isFinite)) throw new RuntimeError("ai_response_invalid", "AI 视觉模型缺少目标坐标");
-  if (!Number.isFinite(width) || width <= 0) width = 24;
-  if (!Number.isFinite(height) || height <= 0) height = 18;
-  width = Math.max(4, Math.min(cropW, width));
-  height = Math.max(4, Math.min(cropH, height));
-  const localX = Math.max(0, Math.min(cropW - width, x1));
-  const localY = Math.max(0, Math.min(cropH - height, y1));
-  const confidence = Number(raw.confidence);
-  return {
-    x: localX,
-    y: localY,
-    width,
-    height,
-    name: String(raw.name ?? "AI识别控件"),
-    controlType: String(raw.controlType ?? "Control"),
-    confidence: Number.isFinite(confidence) ? Math.max(0, Math.min(1, confidence)) : 0.7
-  };
-}
-
-async function runAiPick(context: UiaPickOutput = {}): Promise<DesktopPickResult> {
-  requireAiConfig({ vision: true });
-  const { win, display } = await createAiOverlay();
-  try {
-    let point: { x: number; y: number } | null = null;
-    let physicalX = 0;
-    let physicalY = 0;
-    const autoAi = context.aiRequested && Number.isFinite(Number(context.cursorX)) && Number.isFinite(Number(context.cursorY));
-    if (autoAi) {
-      const cursor = screen.getCursorScreenPoint();
-      point = { x: cursor.x - display.bounds.x, y: cursor.y - display.bounds.y };
-      physicalX = Math.round(Number(context.cursorX));
-      physicalY = Math.round(Number(context.cursorY));
-      await win.webContents.executeJavaScript(
-        "var p=document.getElementById('point');p.style.left=(window.__autoX=" + point.x + ")-23+'px';p.style.top=(window.__autoY=" + point.y + ")-23+'px';p.style.display='block';document.getElementById('tip').textContent='AI 视觉识别中…自上而下扫描画面';document.body.classList.add('scanning');window.__done=true;"
-      );
-      await new Promise(resolve => setTimeout(resolve, 120));
-    }
-    const deadline = Date.now() + 120000;
-    while (!autoAi && Date.now() < deadline) {
-      await new Promise(resolve => setTimeout(resolve, 80));
-      try {
-        const state = await win.webContents.executeJavaScript(
-          "JSON.stringify({done:window.__done===true,cancelled:window.__cancelled===true,point:window.__point||null})", true
-        );
-        const parsed = JSON.parse(state) as { done: boolean; cancelled: boolean; point: { x: number; y: number } | null };
-        if (parsed.cancelled) throw new RuntimeError("pick_cancelled", "已取消 AI 拾取");
-        if (parsed.done && parsed.point) {
-          point = parsed.point;
-          break;
-        }
-      } catch (error) {
-        if (error instanceof RuntimeError && error.code === "pick_cancelled") throw error;
-      }
-    }
-    if (!point) throw new RuntimeError("pick_cancelled", "AI 拾取超时");
-
-    // Hide the effect layer so the screen capture contains the real application only.
-    win.hide();
-    await new Promise(resolve => setTimeout(resolve, 180));
-    if (!autoAi) {
-      const scale = display.scaleFactor || 1;
-      physicalX = Math.round((display.bounds.x + point.x) * scale);
-      physicalY = Math.round((display.bounds.y + point.y) * scale);
-    }
-    const capture = await captureAiAtPoint(physicalX, physicalY);
-    win.showInactive();
-    await win.webContents.executeJavaScript(
-      "document.getElementById('tip').textContent='AI 视觉识别中…请稍候';document.body.classList.add('scanning');"
-    );
-
-    const image = await fs.promises.readFile(capture.imagePath);
-    let aiSettled = false;
-    const aiRequest = chatCompletion([
-      {
-        role: "system",
-        content: "你是桌面 RPA 视觉定位器。只输出 JSON，不输出解释或 Markdown。"
-      },
-      {
-        role: "user",
-        content: [
-          {
-            type: "text",
-            text: [
-              "截图尺寸 " + capture.cropW + "x" + capture.cropH + " 像素。",
-              "红黑色十字中心位于 (" + capture.markerX + ", " + capture.markerY + ")。",
-              "请定位十字中心处或半径 35 像素内最小的可交互 UI 控件；不要返回整个窗口或大容器。",
-              "只返回 JSON：{\"x\":number,\"y\":number,\"width\":number,\"height\":number,\"name\":string,\"controlType\":string,\"confidence\":number}",
-              "x/y 是控件左上角在截图内的像素坐标，width/height 单位也是像素。"
-            ].join("\n")
-          },
-          { type: "image_url", image_url: { url: "data:image/png;base64," + image.toString("base64") } }
-        ]
-      }
-    ], { vision: true, temperature: 0 }).finally(() => { aiSettled = true; });
-    const cancelWatcher = (async () => {
-      while (!aiSettled) {
-        try {
-          const state = await win.webContents.executeJavaScript("JSON.stringify({cancelled:window.__cancelled===true})", true);
-          if (JSON.parse(state).cancelled) throw new RuntimeError("pick_cancelled", "已取消 AI 拾取");
-        } catch (error) {
-          if (error instanceof RuntimeError && error.code === "pick_cancelled") throw error;
-        }
-        await new Promise(resolve => setTimeout(resolve, 100));
-      }
-    })();
-    try {
-      var result = (await Promise.race([aiRequest, cancelWatcher])) as string;
-    } finally {
-      aiSettled = true;
-    }
-
-    const box = parseAiBox(result, capture.cropW, capture.cropH);
-    const x = capture.cropX + box.x;
-    const y = capture.cropY + box.y;
-    const w = box.width;
-    const h = box.height;
-    const parts = [
-      "screen=" + [x, y, w, h].map(Math.round).join(","),
-      context.window || capture.window ? "window=" + safeSelectorText(context.window || capture.window) : "",
-      "name=" + safeSelectorText(box.name),
-      "type=" + safeSelectorText(box.controlType || "Control"),
-      "ai=1"
-    ].filter(Boolean);
-    return {
-      source: "ai",
-      window: context.window || capture.window,
-      name: box.name,
-      controlType: box.controlType,
-      x,
-      y,
-      w,
-      h,
-      confidence: box.confidence,
-      screenshotPath: capture.imagePath,
-      selector: parts.join(";")
-    };
-  } finally {
-    if (!win.isDestroyed()) win.destroy();
-  }
-}
-
 export async function listVisibleWindows(): Promise<Array<{ title: string; process: string }>> {
   const { stdout } = await run(
     "powershell.exe",
@@ -722,3 +553,7 @@ export async function pickScreenRegion(): Promise<{ x: number; y: number; width:
   await run("powershell.exe", ["-NoProfile", "-NonInteractive", "-ExecutionPolicy", "Bypass", "-Command", psScript], { timeout: 30000, windowsHide: true });
   return { ...region, imagePath };
 }
+
+
+
+
diff --git a/src/main/index.ts b/src/main/index.ts
index 90b9c90..8728b93 100644
--- a/src/main/index.ts
+++ b/src/main/index.ts
@@ -9,7 +9,9 @@ import { capabilityCount, capabilityIds } from "./engine/registry";
 import { closeBrowserSession, startRun, stopRun } from "./run/manager";
 import { fixtureUrl, startFixtureServer, stopFixtureServer } from "./fixture/server";
 import { loadSettings, updateSettings } from "./settings";
+import { closeLocalVisionWorkers, defaultUiModelPath } from "./local-vision/ui-detect-service";
 import { pickDesktopElement, listVisibleWindows, pickScreenRegion } from "./desktop-pick";
+import { detectVisionLabTarget, listVisionLabSources } from "./vision-lab";
 import { logMain } from "./logger";
 
 let mainWindow: BrowserWindow | null = null;
@@ -113,9 +115,12 @@ function registerIpc(): void {
     }
     return session.pickElement();
   });
-  ipcMain.handle("desktop:pick-element", (_event, options?: { ai?: boolean }) => pickDesktopElement(options));
+  ipcMain.handle("desktop:pick-element", () => pickDesktopElement());
   ipcMain.handle("desktop:pick-window-list", () => listVisibleWindows());
   ipcMain.handle("desktop:pick-region", () => pickScreenRegion());
+  ipcMain.handle("vision-lab:sources", () => listVisionLabSources());
+  ipcMain.handle("vision-lab:detect", (_event, request: Parameters<typeof detectVisionLabTarget>[0]) => detectVisionLabTarget(request));
+  ipcMain.handle("vision-lab:default-model", () => defaultUiModelPath());
   ipcMain.handle("flows:list", () => listFlows());
   ipcMain.handle("flows:create", (_event, name?: string) => {
     const flow = createFlow(name);
@@ -212,6 +217,7 @@ if (!gotLock) {
   app.on("before-quit", () => {
     void closeBrowserSession();
     void stopFixtureServer();
+    closeLocalVisionWorkers();
   });
 }
 
diff --git a/src/main/local-vision/ui-detect-service.ts b/src/main/local-vision/ui-detect-service.ts
new file mode 100644
index 0000000..518023f
--- /dev/null
+++ b/src/main/local-vision/ui-detect-service.ts
@@ -0,0 +1,145 @@
+import fs from "node:fs";
+import path from "node:path";
+import { Worker } from "node:worker_threads";
+import { app } from "electron";
+import { RuntimeError } from "../engine/types";
+import type { LocalUiBox } from "./ui-detect-worker";
+
+export interface UiDetectOptions {
+  image: Buffer;
+  width: number;
+  height: number;
+  modelPath?: string;
+  scoreThreshold?: number;
+  iouThreshold?: number;
+}
+
+interface DetectRequest {
+  id: number;
+  image: Buffer;
+  width: number;
+  height: number;
+  scoreThreshold?: number;
+  iouThreshold?: number;
+}
+
+interface DetectResponse {
+  id: number;
+  boxes?: LocalUiBox[];
+  error?: string;
+  elapsedMs?: number;
+}
+
+const workers = new Map<string, Worker>();
+const pending = new Map<number, { resolve: (boxes: LocalUiBox[]) => void; reject: (error: unknown) => void; worker: Worker }>();
+let nextRequestId = 1;
+const MAX_LOADED_WORKERS = 4;
+
+export function localUiModelPath(): string {
+  const candidates = [
+    path.join(process.resourcesPath || "", "models", "ui-icon-detect", "model.onnx"),
+    path.join(app.getAppPath(), "resources", "models", "ui-icon-detect", "model.onnx")
+  ].filter(Boolean);
+  const found = candidates.find(candidate => fs.existsSync(candidate));
+  if (!found) {
+    throw new RuntimeError("local_vision_model_missing", "本地 UI 检测模型缺失", {
+      recommendation: "重新安装应用，或补齐 resources/models/ui-icon-detect/model.onnx"
+    });
+  }
+  return found;
+}
+
+export function defaultUiModelPath(): string {
+  return localUiModelPath();
+}
+
+function resolveModelPath(candidate?: string): string {
+  if (!candidate) return localUiModelPath();
+  const normalized = path.resolve(candidate);
+  if (!fs.existsSync(normalized) || !fs.statSync(normalized).isFile()) {
+    throw new RuntimeError("local_vision_model_missing", "模型文件不存在：" + normalized);
+  }
+  if (path.extname(normalized).toLowerCase() !== ".onnx") {
+    throw new RuntimeError("local_vision_model_invalid", "请选择 ONNX 模型文件");
+  }
+  return normalized;
+}
+
+function createWorker(modelPath: string): Worker {
+  const next = new Worker(path.join(__dirname, "local-vision/ui-detect-worker.js"), {
+    workerData: { modelPath },
+    resourceLimits: { maxOldGenerationSizeMb: 512 }
+  });
+  next.on("message", (message: DetectResponse) => {
+    const entry = pending.get(message.id);
+    if (!entry || entry.worker !== next) return;
+    pending.delete(message.id);
+    if (message.error) {
+      entry.reject(new RuntimeError("local_vision_failed", "本地 UI 检测失败：" + message.error));
+      return;
+    }
+    entry.resolve(message.boxes || []);
+  });
+  next.on("error", error => {
+    for (const [id, entry] of pending) {
+      if (entry.worker === next) {
+        entry.reject(error);
+        pending.delete(id);
+      }
+    }
+    workers.delete(modelPath);
+  });
+  next.on("exit", code => {
+    for (const [id, entry] of pending) {
+      if (entry.worker === next) {
+        entry.reject(new RuntimeError("local_vision_failed", "本地 UI 检测进程已退出：" + code));
+        pending.delete(id);
+      }
+    }
+    if (workers.get(modelPath) === next) workers.delete(modelPath);
+  });
+  return next;
+}
+
+function getWorker(modelPath: string): Worker {
+  const existing = workers.get(modelPath);
+  if (existing) return existing;
+  while (workers.size >= MAX_LOADED_WORKERS) {
+    const oldest = workers.keys().next().value;
+    if (!oldest) break;
+    const worker = workers.get(oldest);
+    workers.delete(oldest);
+    void worker?.terminate();
+  }
+  const created = createWorker(modelPath);
+  workers.set(modelPath, created);
+  return created;
+}
+
+export async function detectUiElements(input: UiDetectOptions): Promise<LocalUiBox[]> {
+  const modelPath = resolveModelPath(input.modelPath);
+  const current = getWorker(modelPath);
+  const nextId = nextRequestId++;
+
+  const request: DetectRequest = {
+    id: nextId,
+    image: input.image,
+    width: input.width,
+    height: input.height,
+    scoreThreshold: input.scoreThreshold,
+    iouThreshold: input.iouThreshold
+  };
+  return new Promise<LocalUiBox[]>((resolve, reject) => {
+    pending.set(nextId, { resolve, reject, worker: current });
+    current.postMessage(request);
+  });
+}
+
+export function closeLocalVisionWorkers(): void {
+  for (const [modelPath, worker] of workers) {
+    workers.delete(modelPath);
+    void worker.terminate();
+  }
+  pending.clear();
+}
+
diff --git a/src/main/local-vision/ui-detect-worker.ts b/src/main/local-vision/ui-detect-worker.ts
new file mode 100644
index 0000000..3c6c369
--- /dev/null
+++ b/src/main/local-vision/ui-detect-worker.ts
@@ -0,0 +1,205 @@
+import { parentPort, workerData } from "node:worker_threads";
+import { InferenceSession, Tensor } from "onnxruntime-node";
+import { createCanvas, loadImage } from "@napi-rs/canvas";
+
+export interface LocalUiBox {
+  x: number;
+  y: number;
+  width: number;
+  height: number;
+  category: string;
+  confidence: number;
+}
+
+interface DetectRequest {
+  id: number;
+  image: Buffer;
+  width: number;
+  height: number;
+  scoreThreshold?: number;
+  iouThreshold?: number;
+}
+
+interface DetectResponse {
+  id: number;
+  boxes?: LocalUiBox[];
+  error?: string;
+  elapsedMs?: number;
+}
+
+type ModelOutput = { output0: Tensor };
+type TensorLike = { dims: number[]; data: Float32Array };
+
+const inputName = "images";
+const outputName = "output0";
+const inputSize = 640;
+const defaultScoreThreshold = 0.25;
+const defaultIouThreshold = 0.42;
+const maxBoxes = 160;
+
+let sessionPromise: Promise<InferenceSession> | null = null;
+
+function loadSession(): Promise<InferenceSession> {
+  if (!sessionPromise) {
+    const modelPath = String((workerData as { modelPath?: string }).modelPath || "");
+    sessionPromise = InferenceSession.create(modelPath, {
+      executionProviders: ["cpu"],
+      graphOptimizationLevel: "all",
+      executionMode: "sequential"
+    });
+  }
+  return sessionPromise;
+}
+
+function iou(a: LocalUiBox, b: LocalUiBox): number {
+  const x1 = Math.max(a.x, b.x);
+  const y1 = Math.max(a.y, b.y);
+  const x2 = Math.min(a.x + a.width, b.x + b.width);
+  const y2 = Math.min(a.y + a.height, b.y + b.height);
+  const intersection = Math.max(0, x2 - x1) * Math.max(0, y2 - y1);
+  if (intersection <= 0) return 0;
+  return intersection / (a.width * a.height + b.width * b.height - intersection);
+}
+
+function nms(boxes: LocalUiBox[], iouThreshold: number): LocalUiBox[] {
+  const sorted = [...boxes].sort((a, b) => b.confidence - a.confidence);
+  const kept: LocalUiBox[] = [];
+  for (const box of sorted) {
+    if (kept.every(item => iou(item, box) <= iouThreshold)) kept.push(box);
+  }
+  return kept;
+}
+
+function tensorAccessor(output: TensorLike, anchorCount: number, channelCount: number, channelFirst: boolean) {
+  return (channel: number, anchor: number): number => channelFirst
+    ? output.data[channel * anchorCount + anchor]
+    : output.data[anchor * channelCount + channel];
+}
+
+function extractBoxes(
+  output: TensorLike,
+  sourceWidth: number,
+  sourceHeight: number,
+  scoreThreshold: number,
+  iouThreshold: number
+): LocalUiBox[] {
+  // YOLO/OmniParser exports are commonly [1,C,N] or [1,N,C]. C=4+classCount.
+  if (output.dims.length < 2) throw new Error(`不支持的模型输出维度：${output.dims.join("x")}`);
+  const [, first, second] = output.dims;
+  const channelFirst = first <= second;
+  const channelCount = channelFirst ? first : second;
+  const anchorCount = channelFirst ? second : first;
+  if (channelCount < 5) throw new Error(`模型输出通道数异常：${channelCount}`);
+  const at = tensorAccessor(output, anchorCount, channelCount, channelFirst);
+  const classCount = channelCount - 4;
+
+  const padded = Math.max(sourceWidth, sourceHeight);
+  const scale = inputSize / padded;
+  const drawnWidth = Math.round(sourceWidth * scale);
+  const drawnHeight = Math.round(sourceHeight * scale);
+  const offsetX = (inputSize - drawnWidth) / 2;
+  const offsetY = (inputSize - drawnHeight) / 2;
+  const minSide = Math.max(8, Math.min(sourceWidth, sourceHeight) * 0.006);
+  const raw: LocalUiBox[] = [];
+
+  for (let index = 0; index < anchorCount; index++) {
+    let confidence = 0;
+    let classIndex = 0;
+    if (classCount === 1) {
+      confidence = at(4, index);
+    } else {
+      for (let candidate = 0; candidate < classCount; candidate++) {
+        const score = at(4 + candidate, index);
+        if (score > confidence) {
+          confidence = score;
+          classIndex = candidate;
+        }
+      }
+    }
+    if (!Number.isFinite(confidence) || confidence < scoreThreshold) continue;
+    const cx = at(0, index);
+    const cy = at(1, index);
+    const width = at(2, index);
+    const height = at(3, index);
+    let x = (cx - width / 2 - offsetX) / scale;
+    let y = (cy - height / 2 - offsetY) / scale;
+    let boxWidth = width / scale;
+    let boxHeight = height / scale;
+    x = Math.max(0, Math.min(sourceWidth - 1, x));
+    y = Math.max(0, Math.min(sourceHeight - 1, y));
+    boxWidth = Math.max(2, Math.min(sourceWidth - x, boxWidth));
+    boxHeight = Math.max(2, Math.min(sourceHeight - y, boxHeight));
+    if (boxWidth < minSide || boxHeight < minSide) continue;
+    const areaRatio = (boxWidth * boxHeight) / (sourceWidth * sourceHeight);
+    if (areaRatio > 0.92) continue;
+    raw.push({
+      x: Math.round(x),
+      y: Math.round(y),
+      width: Math.round(boxWidth),
+      height: Math.round(boxHeight),
+      category: classCount <= 1 ? "控件" : `类别${classIndex + 1}`,
+      confidence: Math.round(Math.max(0, Math.min(1, confidence)) * 1000) / 1000
+    });
+  }
+
+  return nms(raw, iouThreshold)
+    .sort((a, b) => b.confidence - a.confidence || a.y - b.y || a.x - b.x)
+    .slice(0, maxBoxes)
+    .sort((a, b) => a.y - b.y || a.x - b.x);
+}
+
+async function detect(request: DetectRequest): Promise<LocalUiBox[]> {
+  const model = await loadSession();
+  const source = await loadImage(request.image);
+  const width = request.width || source.width;
+  const height = request.height || source.height;
+
+  const canvas = createCanvas(inputSize, inputSize);
+  const context = canvas.getContext("2d");
+  context.fillStyle = "#727272";
+  context.fillRect(0, 0, inputSize, inputSize);
+  context.imageSmoothingEnabled = true;
+  context.imageSmoothingQuality = "high";
+  const scale = Math.min(inputSize / width, inputSize / height);
+  context.drawImage(
+    source,
+    (inputSize - width * scale) / 2,
+    (inputSize - height * scale) / 2,
+    width * scale,
+    height * scale
+  );
+  const pixels = context.getImageData(0, 0, inputSize, inputSize).data;
+  const tensorData = new Float32Array(3 * inputSize * inputSize);
+  const planeSize = inputSize * inputSize;
+  for (let index = 0; index < planeSize; index++) {
+    tensorData[index] = pixels[index * 4] / 255;
+    tensorData[planeSize + index] = pixels[index * 4 + 1] / 255;
+    tensorData[planeSize * 2 + index] = pixels[index * 4 + 2] / 255;
+  }
+
+  const results = await model.run({ [inputName]: new Tensor("float32", tensorData, [1, 3, inputSize, inputSize]) });
+  const output = results[outputName] as unknown as TensorLike;
+  return extractBoxes(
+    output,
+    width,
+    height,
+    request.scoreThreshold ?? defaultScoreThreshold,
+    request.iouThreshold ?? defaultIouThreshold
+  );
+}
+
+parentPort?.on("message", async (request: DetectRequest) => {
+  const started = Date.now();
+  try {
+    const boxes = await detect(request);
+    const response: DetectResponse = { id: request.id, boxes, elapsedMs: Date.now() - started };
+    parentPort?.postMessage(response);
+  } catch (error) {
+    const response: DetectResponse = {
+      id: request.id,
+      error: error instanceof Error ? error.message : String(error),
+      elapsedMs: Date.now() - started
+    };
+    parentPort?.postMessage(response);
+  }
+});
diff --git a/src/main/vision-lab.ts b/src/main/vision-lab.ts
new file mode 100644
index 0000000..819a731
--- /dev/null
+++ b/src/main/vision-lab.ts
@@ -0,0 +1,70 @@
+import { desktopCapturer } from "electron";
+import path from "node:path";
+import type { VisionLabDetectRequest, VisionLabDetection, VisionLabSource } from "../shared/types";
+import { RuntimeError } from "./engine/types";
+import { defaultUiModelPath, detectUiElements } from "./local-vision/ui-detect-service";
+
+function sourceKind(sourceId: string): "window" | "screen" {
+  return sourceId.startsWith("screen:") ? "screen" : "window";
+}
+
+function clamp(value: unknown, fallback: number, min: number, max: number): number {
+  const parsed = Number(value);
+  if (!Number.isFinite(parsed)) return fallback;
+  return Math.min(max, Math.max(min, parsed));
+}
+
+export async function listVisionLabSources(): Promise<VisionLabSource[]> {
+  const sources = await desktopCapturer.getSources({
+    types: ["window", "screen"],
+    thumbnailSize: { width: 360, height: 210 },
+    fetchWindowIcons: true
+  });
+  return sources
+    .filter(source => source.thumbnail.isEmpty() === false)
+    .filter(source => sourceKind(source.id) === "screen" || source.name !== "RPA Studio")
+    .map(source => ({
+      id: source.id,
+      kind: sourceKind(source.id),
+      name: source.name || (sourceKind(source.id) === "screen" ? "屏幕" : "未命名窗口"),
+      thumbnail: source.thumbnail.toDataURL(),
+      icon: source.appIcon?.toDataURL()
+    }));
+}
+
+export async function detectVisionLabTarget(request: VisionLabDetectRequest): Promise<VisionLabDetection> {
+  if (!request?.sourceId) throw new RuntimeError("vision_lab_source_required", "请先选择识别目标");
+  const sourceType = sourceKind(request.sourceId);
+  const captureStarted = Date.now();
+  const sources = await desktopCapturer.getSources({
+    types: [sourceType],
+    thumbnailSize: { width: 1280, height: 800 }
+  });
+  const source = sources.find(item => item.id === request.sourceId);
+  if (!source) throw new RuntimeError("vision_lab_source_missing", "识别目标已失效，请重新选择窗口");
+  if (source.thumbnail.isEmpty()) throw new RuntimeError("vision_lab_capture_failed", "目标窗口没有可捕获的画面");
+
+  const image = source.thumbnail.toPNG();
+  const width = source.thumbnail.getSize().width;
+  const height = source.thumbnail.getSize().height;
+  const detectStarted = Date.now();
+  const boxes = await detectUiElements({
+    image,
+    width,
+    height,
+    modelPath: request.modelPath,
+    scoreThreshold: clamp(request.scoreThreshold, 0.25, 0.01, 0.95),
+    iouThreshold: clamp(request.iouThreshold, 0.42, 0.05, 0.9)
+  });
+  return {
+    image: source.thumbnail.toDataURL(),
+    width,
+    height,
+    boxes,
+    captureMs: detectStarted - captureStarted,
+    detectMs: Date.now() - detectStarted,
+    modelPath: path.resolve(request.modelPath || defaultUiModelPath()),
+    sourceName: source.name || (sourceType === "screen" ? "屏幕" : "未命名窗口")
+  };
+}
+
diff --git a/src/preload/index.ts b/src/preload/index.ts
index 77d0cce..6ba6b75 100644
--- a/src/preload/index.ts
+++ b/src/preload/index.ts
@@ -1,5 +1,5 @@
 import { contextBridge, ipcRenderer } from "electron";
-import type { AppInfo, AppSettings, Catalog, FlowDefinition, FlowSaveOptions, FlowSaveResult, FlowSummary, RunEvent } from "../shared/types";
+import type { AppInfo, AppSettings, Catalog, FlowDefinition, FlowSaveOptions, FlowSaveResult, FlowSummary, RunEvent, VisionLabDetectRequest, VisionLabDetection, VisionLabSource } from "../shared/types";
 
 const api = {
   app: {
@@ -9,7 +9,7 @@ const api = {
     get: (): Promise<Catalog> => ipcRenderer.invoke("catalog:get")
   },
   desktop: {
-    pickElement: (options?: { ai?: boolean }) => ipcRenderer.invoke("desktop:pick-element", options || {}),
+    pickElement: () => ipcRenderer.invoke("desktop:pick-element"),
     pickWindowList: () => ipcRenderer.invoke('desktop:pick-window-list'),
     pickRegion: () => ipcRenderer.invoke('desktop:pick-region')
   },
@@ -32,6 +32,11 @@ const api = {
     pickElement: (options?: { url?: string }): Promise<import("../shared/types").ElementPickResult> =>
       ipcRenderer.invoke("browser:pick-element", options || {})
   },
+  visionLab: {
+    sources: (): Promise<VisionLabSource[]> => ipcRenderer.invoke("vision-lab:sources"),
+    detect: (request: VisionLabDetectRequest): Promise<VisionLabDetection> => ipcRenderer.invoke("vision-lab:detect", request),
+    defaultModel: (): Promise<string> => ipcRenderer.invoke("vision-lab:default-model")
+  },
   settings: {
     get: (): Promise<AppSettings> => ipcRenderer.invoke("settings:get"),
     update: (patch: Partial<AppSettings>): Promise<AppSettings> => ipcRenderer.invoke("settings:update", patch || {})
diff --git a/src/renderer/src/App.tsx b/src/renderer/src/App.tsx
index f3e954e..2c68d8e 100644
--- a/src/renderer/src/App.tsx
+++ b/src/renderer/src/App.tsx
@@ -1,16 +1,17 @@
 import { useCallback, useEffect, useState } from "react";
-import { Bot, CircleAlert, Loader2, ListChecks } from "lucide-react";
+import { Bot, CircleAlert, Loader2, ListChecks, Microscope } from "lucide-react";
 import type { Catalog, FlowDefinition, FlowSummary } from "@shared/types";
 import { useEditorStore } from "./store/editorStore";
 import { useRunStore } from "./store/runStore";
 import { EditorPage } from "./pages/EditorPage";
 import { RunsPage } from "./pages/RunsPage";
 import { SettingsPage } from "./pages/SettingsPage";
+import { VisionLabPage } from "./pages/VisionLabPage";
 import { TasksPage } from "./pages/TasksPage";
 import { railIcons } from "./components/icons";
 import { validateFlow } from "./validation/flowValidation";
 
-type Page = "tasks" | "editor" | "runs" | "settings";
+type Page = "tasks" | "editor" | "runs" | "settings" | "visionLab";
 
 interface ToastState {
   id: number;
@@ -235,6 +236,14 @@ export function App() {
           <railIcons.settings className="icon" />
           设置
         </button>
+        <button
+          type="button"
+          className={`rail-item${page === "visionLab" ? " active" : ""}`}
+          onClick={() => setPage("visionLab")}
+        >
+          <Microscope className="icon" />
+          实验台
+        </button>
         <span className="rail-spacer" />
         <div className="rail-user">
           <span className="rail-avatar">RA</span>
@@ -259,6 +268,7 @@ export function App() {
         {page === "editor" && <EditorPage catalog={catalog} onToast={showToast} onBack={() => { setPage("tasks"); void refreshFlows(); }} />}
         {page === "runs" && <RunsPage onBack={() => { setPage("tasks"); void refreshFlows(); }} />}
         {page === "settings" && <SettingsPage catalog={catalog} />}
+        {page === "visionLab" && <VisionLabPage />}
       </main>
 
       {toast && (
@@ -272,3 +282,4 @@ export function App() {
 }
 
 
+
diff --git a/src/renderer/src/pages/VisionLabPage.tsx b/src/renderer/src/pages/VisionLabPage.tsx
new file mode 100644
index 0000000..27b7148
--- /dev/null
+++ b/src/renderer/src/pages/VisionLabPage.tsx
@@ -0,0 +1,232 @@
+import { useCallback, useEffect, useMemo, useRef, useState } from "react";
+import { FolderOpen, Play, Pause, RefreshCw, Sparkles } from "lucide-react";
+import type { VisionLabDetection, VisionLabSource } from "@shared/types";
+
+function fileName(value: string): string {
+  const normalized = value.replace(/[\\/]+$/, "");
+  const index = Math.max(normalized.lastIndexOf("\\"), normalized.lastIndexOf("/"));
+  return index >= 0 ? normalized.slice(index + 1) : normalized;
+}
+
+export function VisionLabPage() {
+  const [sources, setSources] = useState<VisionLabSource[]>([]);
+  const [sourceId, setSourceId] = useState("");
+  const [defaultModel, setDefaultModel] = useState("");
+  const [customModel, setCustomModel] = useState<string | null>(null);
+  const [score, setScore] = useState(0.25);
+  const [overlap, setOverlap] = useState(0.42);
+  const [intervalMs, setIntervalMs] = useState(600);
+  const [continuous, setContinuous] = useState(false);
+  const [busy, setBusy] = useState(false);
+  const [error, setError] = useState<string | null>(null);
+  const [detection, setDetection] = useState<VisionLabDetection | null>(null);
+  const canvasRef = useRef<HTMLCanvasElement | null>(null);
+
+  const runTokenRef = useRef(0);
+  const selectedSource = sources.find(source => source.id === sourceId);
+  const activeModel = customModel || defaultModel;
+
+  const refreshSources = useCallback(async () => {
+    try {
+      const next = await window.rpa.visionLab.sources();
+      setSources(next);
+      setSourceId(current => next.some(source => source.id === current) ? current : next[0]?.id || "");
+      setError(null);
+    } catch (refreshError) {
+      setError(refreshError instanceof Error ? refreshError.message : String(refreshError));
+    }
+  }, []);
+
+  useEffect(() => {
+    void refreshSources();
+    void window.rpa.visionLab.defaultModel().then(setDefaultModel).catch(error => {
+      setError(error instanceof Error ? error.message : String(error));
+    });
+  }, [refreshSources]);
+
+  const detectOnce = useCallback(async () => {
+    if (!sourceId) return;
+    setBusy(true);
+    try {
+      const result = await window.rpa.visionLab.detect({
+        sourceId,
+        modelPath: customModel || undefined,
+        scoreThreshold: score,
+        iouThreshold: overlap
+      });
+      setDetection(result);
+      setError(null);
+    } catch (detectError) {
+      setError(detectError instanceof Error ? detectError.message : String(detectError));
+    } finally {
+      setBusy(false);
+    }
+  }, [customModel, overlap, score, sourceId]);
+
+  useEffect(() => {
+    if (!continuous || !sourceId) return;
+    const token = ++runTokenRef.current;
+    let alive = true;
+    const tick = async () => {
+      while (alive) {
+        const started = Date.now();
+        try {
+          const result = await window.rpa.visionLab.detect({
+            sourceId,
+            modelPath: customModel || undefined,
+            scoreThreshold: score,
+            iouThreshold: overlap
+          });
+          if (!alive || token !== runTokenRef.current) return;
+          setDetection(result);
+          setError(null);
+        } catch (detectError) {
+          if (!alive || token !== runTokenRef.current) return;
+          setError(detectError instanceof Error ? detectError.message : String(detectError));
+        }
+        const wait = Math.max(80, intervalMs - (Date.now() - started));
+        await new Promise(resolve => setTimeout(resolve, wait));
+      }
+    };
+    void tick();
+    return () => { alive = false; };
+  }, [continuous, customModel, intervalMs, overlap, score, sourceId]);
+
+  useEffect(() => {
+    const canvas = canvasRef.current;
+    if (!canvas || !detection) return;
+    const image = new Image();
+    image.onload = () => {
+      canvas.width = detection.width;
+      canvas.height = detection.height;
+      const context = canvas.getContext("2d");
+      if (!context) return;
+      context.clearRect(0, 0, canvas.width, canvas.height);
+      context.drawImage(image, 0, 0, canvas.width, canvas.height);
+      const scale = Math.max(1, detection.width / 1000);
+      context.lineWidth = 2 * scale;
+      context.strokeStyle = "#22d3ee";
+      context.font = `${Math.round(14 * scale)}px Microsoft YaHei, Arial, sans-serif`;
+      context.textBaseline = "top";
+      for (const box of detection.boxes) {
+        context.strokeRect(box.x, box.y, box.width, box.height);
+        const label = `${box.category} ${Math.round(box.confidence * 100)}%`;
+        const metrics = context.measureText(label);
+        const labelWidth = metrics.width + 8 * scale;
+        const labelHeight = 19 * scale;
+        const labelY = box.y >= labelHeight + 2 * scale ? box.y - labelHeight - 2 * scale : box.y + 2 * scale;
+        context.fillStyle = "#0e7490";
+        context.fillRect(box.x, labelY, labelWidth, labelHeight);
+        context.fillStyle = "#ffffff";
+        context.fillText(label, box.x + 4 * scale, labelY + 3 * scale);
+      }
+    };
+    image.src = detection.image;
+  }, [detection]);
+
+  const chooseModel = async () => {
+    const selected = await window.rpa.dialog.open({
+      title: "选择 ONNX UI 检测模型",
+      filters: [{ name: "ONNX 模型", extensions: ["onnx"] }]
+    });
+    if (selected) setCustomModel(selected);
+  };
+
+  const modelLabel = useMemo(() => customModel ? fileName(customModel) : "内置 OmniParser v2", [customModel]);
+  const boxes = detection?.boxes || [];
+
+  return (
+    <div className="vision-lab">
+      <aside className="panel vision-lab-side">
+        <div className="panel-head">
+          <strong><Sparkles className="icon" /> AI 识别实验台</strong>
+          <span className="panel-subtitle">持续识别 · 选窗口 · 选模型</span>
+        </div>
+
+        <div className="vision-block">
+          <label className="field-label" htmlFor="vision-source">识别目标</label>
+          <div className="vision-row">
+            <select
+              id="vision-source"
+              className="text-input"
+              value={sourceId}
+              onChange={event => setSourceId(event.target.value)}
+            >
+              {!sources.length && <option value="">正在读取窗口…</option>}
+              {sources.map(source => (
+                <option key={source.id} value={source.id}>
+                  {source.kind === "screen" ? "屏幕 · " : "窗口 · "}{source.name}
+                </option>
+              ))}
+            </select>
+            <button className="btn icon-btn" type="button" onClick={() => void refreshSources()} title="刷新窗口列表">
+              <RefreshCw className="icon" />
+            </button>
+          </div>
+          {selectedSource && (
+            <div className="source-preview">
+              <img src={selectedSource.thumbnail} alt={selectedSource.name} />
+            </div>
+          )}
+        </div>
+
+        <div className="vision-block">
+          <span className="field-label">识别模型</span>
+          <div className="vision-row">
+            <button
+              type="button"
+              className={`btn${customModel ? "" : " btn-primary"}`}
+              onClick={() => setCustomModel(null)}
+            >
+              内置 OmniParser v2
+            </button>
+            <button type="button" className={`btn${customModel ? " btn-primary" : ""}`} onClick={() => void chooseModel()}>
+              <FolderOpen className="icon" />
+              选择 ONNX
+            </button>
+          </div>
+          <div className="model-path" title={activeModel || ""}>{activeModel || "模型未找到"}</div>
+        </div>
+
+        <div className="vision-block">
+          <label className="field-label" htmlFor="vision-score">置信度阈值：{score.toFixed(2)}</label>
+          <input id="vision-score" type="range" min={0.05} max={0.9} step={0.01} value={score}
+                 onChange={event => setScore(Number(event.target.value))} />
+          <label className="field-label" htmlFor="vision-overlap">重叠抑制：{overlap.toFixed(2)}</label>
+          <input id="vision-overlap" type="range" min={0.1} max={0.85} step={0.01} value={overlap}
+                 onChange={event => setOverlap(Number(event.target.value))} />
+          <label className="field-label" htmlFor="vision-interval">识别间隔：{intervalMs} ms</label>
+          <input id="vision-interval" type="range" min={100} max={3000} step={50} value={intervalMs}
+                 onChange={event => setIntervalMs(Number(event.target.value))} />
+        </div>
+
+        <div className="vision-actions">
+          <button className="btn" type="button" onClick={() => void detectOnce()} disabled={!sourceId || busy}>
+            识别一次
+          </button>
+          <button className={`btn${continuous ? " btn-danger" : " btn-primary"}`} type="button"
+                  onClick={() => setContinuous(value => !value)} disabled={!sourceId}>
+            {continuous ? <Pause className="icon" /> : <Play className="icon" />}
+            {continuous ? "停止连续" : "开始连续"}
+          </button>
+        </div>
+        {error && <div className="vision-error">{error}</div>}
+      </aside>
+
+      <section className="panel vision-lab-stage">
+        <div className="vision-stage-head">
+          <strong>{detection?.sourceName || selectedSource?.name || "未选择目标"}</strong>
+          <span>{boxes.length} 个候选 · 采集 {detection?.captureMs ?? "-"}ms · 推理 {detection?.detectMs ?? "-"}ms</span>
+        </div>
+        <div className="vision-stage">
+          {detection ? (
+            <canvas ref={canvasRef} className="vision-canvas" />
+          ) : (
+            <div className="vision-empty">选择窗口和模型后开始识别</div>
+          )}
+        </div>
+      </section>
+    </div>
+  );
+}
+
diff --git a/src/renderer/src/panels/Properties.tsx b/src/renderer/src/panels/Properties.tsx
index 2f3a9ec..3403e20 100644
--- a/src/renderer/src/panels/Properties.tsx
+++ b/src/renderer/src/panels/Properties.tsx
@@ -464,10 +464,10 @@ function FieldControl({
     onChange(field.key, current === "" || known ? expression : `${textOf(raw)}${expression}`);
   };
 
-  const pickDesktop = async (ai = false) => {
+  const pickDesktop = async () => {
     setPicking(true);
     try {
-      const picked = await window.rpa.desktop.pickElement({ ai });
+      const picked = await window.rpa.desktop.pickElement();
       if (picked.cancelled) {
         onToast("已取消拾取");
         return;
@@ -647,12 +647,9 @@ function FieldControl({
           />
           {isGroundTarget ? (
             <>
-              <button type="button" className="mini-btn" disabled={picking} onClick={() => void pickDesktop(false)} title="从桌面拾取元素（移到目标按空格）">
+              <button type="button" className="mini-btn" disabled={picking} onClick={() => void pickDesktop()} title="从桌面拾取元素（移到目标按空格）">
                 {picking ? "桌面拾取中" : "桌面拾取"}
               </button>
-              <button type="button" className="mini-btn" disabled={picking} onClick={() => void pickDesktop(true)} title="AI 视觉拾取；UIA 类型为 Pane 时也可在桌面拾取中按 F2">
-                {picking ? "AI识别中" : "AI拾取"}
-              </button>
             </>
           ) : (
             <button type="button" className="mini-btn" disabled={picking} onClick={() => void pickElement()} title="从运行浏览器拾取元素">
diff --git a/src/renderer/src/styles/app.css b/src/renderer/src/styles/app.css
index da2d529..8767277 100644
--- a/src/renderer/src/styles/app.css
+++ b/src/renderer/src/styles/app.css
@@ -3444,3 +3444,187 @@ textarea:focus-visible {
 .icon-btn.refresh-btn.refreshing .icon {
   animation: spin 0.8s linear infinite;
 }
+
+.vision-lab {
+  display: grid;
+  grid-template-columns: 340px minmax(0, 1fr);
+  gap: 18px;
+  height: 100%;
+  min-height: 0;
+  padding: 18px;
+}
+
+.vision-lab-side {
+  display: flex;
+  flex-direction: column;
+  gap: 16px;
+  min-height: 0;
+  overflow: auto;
+  padding: 18px;
+}
+
+.vision-lab-side .panel-head {
+  align-items: center;
+  display: flex;
+  flex-direction: column;
+  gap: 4px;
+}
+
+.vision-lab-side .panel-head strong {
+  align-items: center;
+  display: flex;
+  font-size: 16px;
+  gap: 8px;
+}
+
+.vision-block {
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+}
+
+.vision-row {
+  align-items: center;
+  display: flex;
+  gap: 8px;
+}
+
+.vision-row select {
+  min-width: 0;
+  flex: 1;
+}
+
+.vision-row .btn {
+  flex: none;
+}
+
+.icon-btn {
+  padding-left: 10px;
+  padding-right: 10px;
+}
+
+.source-preview {
+  border: 1px solid #d8e0ea;
+  border-radius: 10px;
+  overflow: hidden;
+}
+
+.source-preview img {
+  display: block;
+  height: 116px;
+  object-fit: cover;
+  width: 100%;
+}
+
+.model-path {
+  border-radius: 8px;
+  background: #f3f6fa;
+  color: #64748b;
+  font-size: 12px;
+  line-height: 1.4;
+  max-height: 52px;
+  overflow: auto;
+  padding: 7px 9px;
+  word-break: break-all;
+}
+
+.vision-actions {
+  display: grid;
+  gap: 8px;
+  grid-template-columns: 1fr 1fr;
+  margin-top: auto;
+}
+
+.vision-error {
+  background: #fff1f1;
+  border: 1px solid #fecaca;
+  border-radius: 8px;
+  color: #b91c1c;
+  font-size: 12px;
+  line-height: 1.45;
+  padding: 9px 10px;
+  white-space: pre-wrap;
+}
+
+.vision-lab-stage {
+  display: flex;
+  flex-direction: column;
+  min-height: 0;
+  min-width: 0;
+  overflow: hidden;
+  padding: 18px;
+}
+
+.vision-stage-head {
+  align-items: center;
+  display: flex;
+  gap: 12px;
+  justify-content: space-between;
+  margin-bottom: 12px;
+}
+
+.vision-stage-head strong {
+  font-size: 15px;
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+
+.vision-stage-head span {
+  color: #64748b;
+  flex: none;
+  font-size: 12px;
+}
+
+.vision-stage {
+  align-items: center;
+  background: #102033;
+  border-radius: 12px;
+  display: flex;
+  flex: 1;
+  justify-content: center;
+  min-height: 0;
+  overflow: hidden;
+  position: relative;
+}
+
+.vision-image {
+  display: block;
+  height: 100%;
+  object-fit: contain;
+  width: 100%;
+}
+
+.vision-box {
+  border: 2px solid #22d3ee;
+  box-shadow: 0 0 0 1px rgba(0, 0, 0, .35) inset;
+  position: absolute;
+}
+
+.vision-box span {
+  background: #0e7490;
+  color: white;
+  display: inline-block;
+  font-size: 11px;
+  line-height: 17px;
+  padding: 0 5px;
+  transform: translateY(-100%);
+  white-space: nowrap;
+}
+
+.vision-empty {
+  align-items: center;
+  color: #94a3b8;
+  display: flex;
+  height: 100%;
+  justify-content: center;
+}
+
+.vision-canvas {
+  display: block;
+  height: auto;
+  max-height: 100%;
+  max-width: 100%;
+  width: auto;
+}
+
diff --git a/src/shared/types.ts b/src/shared/types.ts
index 06cdf6d..b41a68b 100644
--- a/src/shared/types.ts
+++ b/src/shared/types.ts
@@ -247,3 +247,38 @@ export interface RunEvent {
   startedAt?: string;
 }
 
+
+export interface VisionLabSource {
+  id: string;
+  kind: "window" | "screen";
+  name: string;
+  thumbnail: string;
+  icon?: string;
+}
+
+export interface VisionLabBox {
+  x: number;
+  y: number;
+  width: number;
+  height: number;
+  category: string;
+  confidence: number;
+}
+
+export interface VisionLabDetectRequest {
+  sourceId: string;
+  modelPath?: string;
+  scoreThreshold?: number;
+  iouThreshold?: number;
+}
+
+export interface VisionLabDetection {
+  image: string;
+  width: number;
+  height: number;
+  boxes: VisionLabBox[];
+  captureMs: number;
+  detectMs: number;
+  modelPath: string;
+  sourceName: string;
+}