feat(computer): improve Linux headless display quality (#88)

* feat(computer): improve Linux headless display quality

- Increase Xvfb resolution from default 1920x1080 to 2560x1440x24
- Add fluxbox window manager for proper window decorations
- Install GTK themes (adwaita, gnome-themes-extra) for native UI styling
- Add Noto CJK fonts for proper CJK character rendering
- Add dbus-x11 for GTK application support

* ci: add branch trigger for PR visibility

* ci: trigger workflow on pull requests targeting main

* fix(computer): maximize Obsidian window on launch

* fix(computer): reduce flakiness in community plugin installation steps

- Add aiWaitFor verification after each navigation step to catch failures early
- Split conditional "Turn on" action and confirmation dialog into separate steps
- Wait for "Browse" button to be visible before clicking it
- Increase sleep times between UI interactions
- Set MIDSCENE_REPLANNING_CYCLE_LIMIT=40 as safety net

* fix(computer): maximize window via xdotool instead of --start-maximized

--start-maximized Electron flag doesn't work with fluxbox. Use xdotool
to find the Obsidian window and resize it to fill the entire screen.

* fix(computer): maximize window by double-clicking title bar via aiAct

* fix(computer): click maximize button instead of double-clicking title bar

* fix(computer): configure fluxbox to auto-maximize all windows

Write ~/.fluxbox/apps config before starting fluxbox so all windows
launch maximized automatically. This eliminates the need to manually
click maximize buttons or double-click title bars.

Author: quanru

.github/workflows/computer-electron-demo.yaml

@@ -14,10 +14,15 @@ on:
         description: 'The branch to checkout'
         required: false
         default: 'main'
+  pull_request:
+    branches:
+      - main
+    paths:
+      - 'computer/electron-demo/**'
+      - '.github/workflows/computer-electron-demo.yaml'
   push:
     branches:
       - main
-      - feat/computer-electron-demo
     paths:
       - 'computer/electron-demo/**'
       - '.github/workflows/computer-electron-demo.yaml'
@@ -34,6 +39,7 @@ jobs:
       MIDSCENE_MODEL_NAME: 'qwen-vl-max-latest'
       MIDSCENE_USE_QWEN_VL: 1
       MIDSCENE_COMPUTER_HEADLESS_LINUX: 'true'
+      MIDSCENE_REPLANNING_CYCLE_LIMIT: '40'
       DEBUG: 'midscene:ai:profile:*'
 
     steps:
@@ -63,7 +69,12 @@ jobs:
             libnotify4 \
             libsecret-1-0 \
             libxss1 \
-            xdg-utils
+            xdg-utils \
+            fluxbox \
+            adwaita-icon-theme \
+            gnome-themes-extra \
+            fonts-noto-cjk \
+            dbus-x11
 
       - name: Download Obsidian AppImage
         working-directory: computer/electron-demo

computer/electron-demo/demo.ts

@@ -103,8 +103,27 @@ function launchApp(
     aiActionContext:
       'You are interacting with Obsidian, a note-taking desktop application. ' +
       'If any dialog or popup appears, dismiss it by clicking the close button or pressing Escape.',
+    xvfbResolution: '2560x1440x24',
   });
 
+  // --- Start window manager for proper window decorations on headless Linux ---
+  if (process.platform === 'linux' && process.env.DISPLAY) {
+    // Configure fluxbox to auto-maximize all windows
+    const fluxboxDir = join(homedir(), '.fluxbox');
+    mkdirSync(fluxboxDir, { recursive: true });
+    writeFileSync(
+      join(fluxboxDir, 'apps'),
+      '[app] (name=.*)\n  [Maximized] {yes}\n[end]\n',
+    );
+    const fluxbox = spawn('fluxbox', [], {
+      detached: true,
+      stdio: 'ignore',
+    });
+    fluxbox.unref();
+    console.log('Fluxbox window manager started (auto-maximize enabled)');
+    await sleep(1000);
+  }
+
   // --- Launch Obsidian AFTER Xvfb is ready (DISPLAY is now set) ---
   const child = launchApp(binaryPath, VAULT_DIR);
   console.log(`Obsidian launched (pid: ${child.pid})`);
@@ -152,35 +171,54 @@ function launchApp(
     // --- Install community plugin: LifeOS ---
     console.log('Opening Obsidian settings...');
     await agent.aiAct('press Ctrl+Comma to open Settings');
-    await sleep(2000);
+    await sleep(3000);
 
-    await agent.aiWaitFor('Settings dialog is visible', { timeoutMs: 15000 });
+    await agent.aiWaitFor(
+      'A Settings dialog/panel is visible with a left sidebar containing menu items',
+      { timeoutMs: 15000 },
+    );
+    console.log('Settings dialog is open');
 
     // Navigate to Community plugins
     await agent.aiAct(
-      'click "Community plugins" in the left sidebar of the settings dialog',
+      'In the Settings dialog, click "Community plugins" in the left sidebar',
     );
-    await sleep(1500);
+    await sleep(2000);
+
+    await agent.aiWaitFor(
+      'The Community plugins settings page is visible',
+      { timeoutMs: 10000 },
+    );
+    console.log('Community plugins page is visible');
 
     // Turn on community plugins if not enabled
     await agent.aiAct(
-      'If there is a "Turn on community plugins" button, click it. ' +
-        'If a confirmation dialog appears, click "Turn on" to confirm.',
+      'If there is a "Turn on community plugins" button, click it',
     );
-    await sleep(1500);
+    await sleep(2000);
+
+    // Handle confirmation dialog separately
+    await agent.aiAct(
+      'If a confirmation dialog is visible asking to turn on community plugins, click the "Turn on" button to confirm. Otherwise do nothing.',
+    );
+    await sleep(2000);
 
     // Open the plugin browser
-    await agent.aiAct('click "Browse" button to open the community plugin browser');
+    await agent.aiWaitFor(
+      'A "Browse" button is visible on the Community plugins page',
+      { timeoutMs: 10000 },
+    );
+    await agent.aiAct('click the "Browse" button');
     await sleep(3000);
 
     await agent.aiWaitFor(
-      'Community plugin browser / marketplace is visible with a search box',
+      'Community plugin browser is visible with a search box',
       { timeoutMs: 20000 },
     );
     console.log('Community plugin browser is open');
 
     // Search for LifeOS
-    await agent.aiAct('type "lifeos" in the search box');
+    await agent.aiAct('click the search box and type "lifeos"');
     await sleep(3000);
 
     // Click the LifeOS plugin from results