omni-vu（全方位视觉）

omni.vu - Visual Understanding & Automation

Overview

omni.vu gives you eyes and hands on the user's macOS screen. Use it to:

• See what the user sees (screen capture)
• Understand UI state with AI vision
• Detect changes and wait for events
• Act with mouse and keyboard automation

When to Use This Skill

Proactively Use omni.vu When:

1. Debugging UI issues - "The button isn't working" → capture and analyze
2. Verifying changes - After modifying UI code, check if it rendered correctly
3. Waiting for operations - Build finishing, deployment completing, tests running
4. Understanding context - User describes something on screen you can't see
5. Automating repetitive tasks - Clicking through UI flows, filling forms
6. Documentation - Capturing screenshots of features

Do NOT Use When:

• Reading/writing files (use Read/Write tools)
• Running terminal commands (use Bash)
• Making API calls (use appropriate tools)
• User hasn't granted screen recording permission

Tool Reference

Capture Tools

vu - Full Screen Capture

vu(monitor=0, save_to_history=True)

Captures the entire screen. Returns base64 image.

Use when: You need to see everything on screen.

vu_window - Window Capture

vu_window(window_id=None, title="VS Code", include_frame=False)

Captures a specific window by ID or title (partial match).

Use when: You only need one application's content.

Workflow:

1. Call vu_list_windows to see available windows
2. Find the window_id or use title matching
3. Call vu_window with that ID/title

vu_region - Region Capture

vu_region(x=100, y=100, width=500, height=300)

Captures a specific rectangle of the screen.

Use when: You need a precise area (error message, specific component).

vu_list_windows - List Windows

vu_list_windows(filter_app="Chrome")

Lists all visible windows with metadata.

Returns: List of {window_id, title, owner, x, y, width, height}

vu_list_monitors - List Displays

vu_list_monitors()

Lists connected displays with resolution and scale factor.

Vision Tools

vu_describe - AI Vision Analysis

vu_describe(
    prompt="What errors are visible?",
    provider="claude",  # or openai, gemini, ollama
    max_tokens=1024,
    capture_first=True
)

Captures screen and analyzes with AI.

Best prompts:

• "Describe what you see on screen"
• "Are there any error messages visible?"
• "What is the state of the build/test output?"
• "Is the login form filled correctly?"
• "What color is the status indicator?"

Providers:

Utility Tools

vu_diff - Change Detection

vu_diff(threshold=0.02, monitor=0)

Detects if screen changed since last capture.

Returns: {changed: bool, diff_percentage: float}

Use for polling patterns:

# Wait for build to finish
while True:
    result = vu_diff(threshold=0.05)
    if result["changed"]:
        # Screen changed, check what happened
        analysis = vu_describe(prompt="Did the build succeed or fail?")
        break
    # Wait before checking again

vu_history - Capture History

vu_history(limit=10, capture_type="full_screen")

Gets recent capture metadata.

vu_last - Last Capture

vu_last()

Returns the most recent capture with image data.

Use when: You need to re-analyze without re-capturing.

vu_status - System Status

vu_status()

Returns system info: monitors, providers, safety settings.

Automation Tools

vu_click - Mouse Click

vu_click(x=500, y=300, button="left", count=1)

Clicks at screen coordinates.

Buttons: left, right, middle Count: 1 (single), 2 (double), 3 (triple)

Safety: Coordinates validated against screen bounds.

vu_move - Move Cursor

vu_move(x=500, y=300, duration_ms=0)

Moves cursor to position. Use duration_ms for animated movement.

vu_drag - Drag Operation

vu_drag(start_x=100, start_y=100, end_x=500, end_y=300, duration_ms=500)

Drags from start to end position.

Safety: Maximum 2000px drag distance.

vu_scroll - Scroll

vu_scroll(direction="down", amount=3, x=None, y=None)

Scrolls at current or specified position.

Directions: up, down, left, right Amount: Lines to scroll (1-20)

vu_type - Type Text

vu_type(text="Hello World", delay_between_ms=0)

Types text character by character.

Note: Click on target field first!

vu_hotkey - Keyboard Shortcut

vu_hotkey(keys="cmd+s")

Executes keyboard shortcuts.

Format: modifier+key (e.g., cmd+c, ctrl+shift+s, cmd+opt+i) Modifiers: cmd, ctrl, alt/opt, shift

Blocked hotkeys (for safety):

• cmd+q (quit)
• cmd+shift+q (logout)
• cmd+opt+esc (force quit)

Common Workflows

1. Debug UI Issue

User: "The submit button doesn't do anything when I click it"

1. vu_describe(prompt="Describe the form and submit button state")
2. Analyze: Is button disabled? Is there validation error?
3. If needed: vu_window(title="Chrome") for focused capture
4. Check console: vu_describe(prompt="Are there any errors in the developer console?")

2. Verify Build/Deploy

User: "Run the build and let me know when it's done"

1. Run: npm run build (via Bash)
2. Loop: vu_diff(threshold=0.05)
3. When changed: vu_describe(prompt="What is the build status? Did it succeed or fail?")
4. Report result

3. Fill a Form

User: "Fill out the registration form with test data"

1. vu_describe(prompt="What form fields are visible?")
2. vu_click(x=field_x, y=field_y)  # Click first field
3. vu_type(text="testuser@example.com")
4. vu_hotkey(keys="tab")  # Move to next field
5. vu_type(text="Test User")
6. Continue for each field...
7. vu_click(x=submit_x, y=submit_y)  # Submit
8. vu_describe(prompt="Was the form submitted successfully?")

4. Navigate UI

User: "Open the settings panel in VS Code"

1. vu_list_windows(filter_app="Code")
2. vu_window(title="Code")  # Capture VS Code
3. vu_hotkey(keys="cmd+,")  # Open settings
4. vu_diff()  # Wait for settings to open
5. vu_describe(prompt="Are the VS Code settings now visible?")

5. Monitor for Changes

User: "Watch the dashboard and tell me when the status changes"

1. vu()  # Initial capture
2. Loop every 5 seconds:
   - vu_diff(threshold=0.02)
   - If changed: vu_describe(prompt="What changed on the dashboard?")
3. Report changes to user

Best Practices

1. Capture Before Acting

Always capture and analyze before clicking/typing:

❌ Bad: vu_click(x=500, y=300)  # Hope it's the right spot

✅ Good:
1. vu_describe(prompt="Where is the submit button?")
2. Parse coordinates from description
3. vu_click(x=parsed_x, y=parsed_y)

2. Use Appropriate Capture Scope

Full screen → General context, multi-app workflows
Window → Single app focus, cleaner output
Region → Specific element, faster/smaller

3. Specific Vision Prompts

❌ Vague: "What do you see?"

✅ Specific:
- "Is there an error message visible? If so, what does it say?"
- "What is the status of the test runner? Passing, failing, or running?"
- "List all form fields visible and their current values"

4. Rate Limit Automation

Don't spam clicks. Wait between actions:

vu_click(x=100, y=100)
# Wait for UI response
vu_diff(threshold=0.01)
vu_click(x=200, y=200)

5. Verify After Actions

Always verify automation succeeded:

vu_type(text="hello@example.com")
vu_describe(prompt="What text is now in the email field?")

Troubleshooting

"Permission denied" or blank captures

→ Grant Screen Recording permission: System Settings > Privacy > Screen Recording

Automation not working

→ Grant Accessibility permission: System Settings > Privacy > Accessibility

Vision analysis failing

→ Check API keys in environment variables → Try different provider: vu_describe(provider="openai")

Coordinates seem off

→ Retina display? Coordinates are in logical points, not pixels → Multi-monitor? Check vu_list_monitors() for display arrangement

Hotkey blocked

→ Some dangerous hotkeys (cmd+q) are blocked for safety → Unblock in config if absolutely necessary

Configuration

Environment variables:

OMNI_VU_ANTHROPIC_API_KEY=sk-ant-...    # For Claude vision
OMNI_VU_OPENAI_API_KEY=sk-...           # For GPT-4 vision
OMNI_VU_DEFAULT_PROVIDER=claude          # Default AI provider
OMNI_VU_SAFETY_LEVEL=medium              # low/medium/high

History stored at: ~/.omni.vu/captures/