@elizaos/computeruse 0.24.20

package/src/desktop.rs

@@ -0,0 +1,2763 @@
+use crate::types::{
+    ClickResult, ClickType, ComputerUseResult, ComputerUseStep, Monitor, MonitorScreenshotPair,
+    ResizedDimensions, TreeOutputFormat, VisionType, WindowTreeResult,
+};
+use crate::Selector;
+use crate::{
+    map_error, CommandOutput, Element, Locator, ScreenshotResult, TreeBuildConfig, UINode,
+};
+use napi::bindgen_prelude::Either;
+use napi::threadsafe_function::{ThreadsafeFunction, ThreadsafeFunctionCallMode};
+use napi_derive::napi;
+use std::sync::{Arc, Once};
+use computeruse::Desktop as ComputerUseDesktop;
+
+/// Normalize key format to ensure curly brace syntax for special keys.
+/// If key already contains `{`, assume it's correctly formatted.
+/// Otherwise, wrap the entire key in `{}` to ensure it's treated as a special key press.
+fn normalize_key(key: &str) -> String {
+    if key.contains('{') {
+        key.to_string()
+    } else {
+        format!("{{{}}}", key)
+    }
+}
+
+/// Result of screenshot capture operations
+#[derive(Default)]
+struct ScreenshotPaths {
+    window_path: Option<String>,
+    monitor_paths: Option<Vec<String>>,
+}
+
+/// Helper to capture and save screenshots based on options
+fn capture_screenshots(
+    desktop: &ComputerUseDesktop,
+    pid: Option<u32>,
+    include_window: bool,
+    include_monitors: bool,
+    operation: &str,
+) -> ScreenshotPaths {
+    let mut result = ScreenshotPaths::default();
+
+    if !include_window && !include_monitors {
+        return result;
+    }
+
+    computeruse::screenshot_logger::init();
+    let prefix = computeruse::screenshot_logger::generate_prefix(None, operation);
+
+    if include_window {
+        if let Some(pid) = pid {
+            // Try to capture window screenshot by PID
+            if let Ok(apps) = desktop.applications() {
+                if let Some(app) = apps.into_iter().find(|a| a.process_id().ok() == Some(pid)) {
+                    if let Ok(screenshot) = app.capture() {
+                        if let Some(saved) = computeruse::screenshot_logger::save_window_screenshot(
+                            &screenshot,
+                            &prefix,
+                            None,
+                        ) {
+                            result.window_path = Some(saved.path.to_string_lossy().to_string());
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    if include_monitors {
+        // Capture all monitors using futures executor for sync context
+        if let Ok(monitors) = futures::executor::block_on(desktop.capture_all_monitors()) {
+            let saved =
+                computeruse::screenshot_logger::save_monitor_screenshots(&monitors, &prefix, None);
+            if !saved.is_empty() {
+                result.monitor_paths = Some(
+                    saved
+                        .into_iter()
+                        .map(|s| s.path.to_string_lossy().to_string())
+                        .collect(),
+                );
+            }
+        }
+    }
+
+    result
+}
+
+/// Helper to find PID from process name using the shared core function.
+fn find_pid_for_process(desktop: &ComputerUseDesktop, process_name: &str) -> napi::Result<u32> {
+    computeruse::find_pid_for_process(desktop, process_name).map_err(map_error)
+}
+
+/// Main entry point for desktop automation.
+#[napi(js_name = "Desktop")]
+pub struct Desktop {
+    inner: ComputerUseDesktop,
+}
+
+#[allow(clippy::needless_pass_by_value)]
+#[napi]
+impl Desktop {
+    /// Create a new Desktop automation instance with configurable options.
+    ///
+    /// @param {boolean} [useBackgroundApps=false] - Enable background apps support.
+    /// @param {boolean} [activateApp=false] - Enable app activation support.
+    /// @param {string} [logLevel] - Logging level (e.g., 'info', 'debug', 'warn', 'error').
+    ///                              Falls back to RUST_LOG or COMPUTERUSE_LOG_LEVEL env vars, defaults to 'info'.
+    /// @returns {Desktop} A new Desktop automation instance.
+    #[napi(constructor)]
+    pub fn new(
+        use_background_apps: Option<bool>,
+        activate_app: Option<bool>,
+        log_level: Option<String>,
+    ) -> Self {
+        let use_background_apps = use_background_apps.unwrap_or(false);
+        let activate_app = activate_app.unwrap_or(false);
+
+        // Priority: explicit param > RUST_LOG env > COMPUTERUSE_LOG_LEVEL env > "info" default
+        let log_level = log_level
+            .or_else(|| std::env::var("RUST_LOG").ok())
+            .or_else(|| std::env::var("COMPUTERUSE_LOG_LEVEL").ok())
+            .unwrap_or_else(|| "info".to_string());
+
+        static INIT: Once = Once::new();
+        INIT.call_once(|| {
+            let _ = tracing_subscriber::fmt()
+                .with_env_filter(log_level)
+                .with_ansi(false) // Disable ANSI color codes for cleaner output
+                .try_init();
+        });
+        let desktop = ComputerUseDesktop::new(use_background_apps, activate_app)
+            .expect("Failed to create Desktop instance");
+        Desktop { inner: desktop }
+    }
+
+    /// Get the root UI element of the desktop.
+    ///
+    /// @returns {Element} The root UI element.
+    #[napi]
+    pub fn root(&self) -> Element {
+        let root = self.inner.root();
+        Element::from(root)
+    }
+
+    /// Get a list of all running applications.
+    ///
+    /// @returns {Array<Element>} List of application UI elements.
+    #[napi]
+    pub fn applications(&self) -> napi::Result<Vec<Element>> {
+        self.inner
+            .applications()
+            .map(|apps| apps.into_iter().map(Element::from).collect())
+            .map_err(map_error)
+    }
+
+    /// Get a running application by name.
+    ///
+    /// @param {string} name - The name of the application to find.
+    /// @returns {Element} The application UI element.
+    #[napi]
+    pub fn application(&self, name: String) -> napi::Result<Element> {
+        self.inner
+            .application(&name)
+            .map(Element::from)
+            .map_err(map_error)
+    }
+
+    /// Open an application by name.
+    ///
+    /// @param {string} name - The name of the application to open.
+    /// @param {boolean} [includeWindowScreenshot=true] - Whether to capture window screenshot after opening
+    /// @param {boolean} [includeMonitorScreenshots=false] - Whether to capture monitor screenshots after opening
+    #[napi]
+    pub fn open_application(
+        &self,
+        name: String,
+        include_window_screenshot: Option<bool>,
+        include_monitor_screenshots: Option<bool>,
+    ) -> napi::Result<Element> {
+        let element = self.inner.open_application(&name).map_err(map_error)?;
+
+        // Capture screenshots if enabled (window default: true, monitor default: false)
+        let _screenshots = capture_screenshots(
+            &self.inner,
+            element.process_id().ok(),
+            include_window_screenshot.unwrap_or(true),
+            include_monitor_screenshots.unwrap_or(false),
+            "openApplication",
+        );
+
+        Ok(Element::from(element))
+    }
+
+    /// Activate an application by name.
+    ///
+    /// @param {string} name - The name of the application to activate.
+    #[napi]
+    pub fn activate_application(&self, name: String) -> napi::Result<()> {
+        self.inner.activate_application(&name).map_err(map_error)
+    }
+
+    /// Click within element bounds at a specified position.
+    ///
+    /// This is useful for clicking on elements from UI tree, OCR, omniparser, gemini vision, or DOM
+    /// without needing an element reference - just the bounds.
+    ///
+    /// @param {number} x - X coordinate of the bounds.
+    /// @param {number} y - Y coordinate of the bounds.
+    /// @param {number} width - Width of the bounds.
+    /// @param {number} height - Height of the bounds.
+    /// @param {number} [xPercentage=50] - X position within bounds as percentage (0-100). Defaults to 50 (center).
+    /// @param {number} [yPercentage=50] - Y position within bounds as percentage (0-100). Defaults to 50 (center).
+    /// @param {ClickType} [clickType='left'] - Type of click: 'left', 'double', or 'right'.
+    /// @param {boolean} [restoreCursor=true] - If true, restore cursor to original position after clicking.
+    /// @param {string} [process] - Process name for window screenshot capture. If provided, enables window screenshots.
+    /// @param {boolean} [includeWindowScreenshot=true] - Whether to capture window screenshot (requires process).
+    /// @param {boolean} [includeMonitorScreenshots=false] - Whether to capture monitor screenshots after clicking.
+    /// @returns {ClickResult} Result with clicked coordinates and method details.
+    #[napi]
+    #[allow(clippy::too_many_arguments)]
+    pub fn click_at_bounds(
+        &self,
+        x: f64,
+        y: f64,
+        width: f64,
+        height: f64,
+        x_percentage: Option<u8>,
+        y_percentage: Option<u8>,
+        click_type: Option<ClickType>,
+        restore_cursor: Option<bool>,
+        process: Option<String>,
+        include_window_screenshot: Option<bool>,
+        include_monitor_screenshots: Option<bool>,
+    ) -> napi::Result<ClickResult> {
+        let bounds = (x, y, width, height);
+        let click_position = match (x_percentage, y_percentage) {
+            (Some(xp), Some(yp)) => Some((xp, yp)),
+            (Some(xp), None) => Some((xp, 50)),
+            (None, Some(yp)) => Some((50, yp)),
+            (None, None) => None,
+        };
+        let click_type = click_type.unwrap_or(ClickType::Left);
+        let restore_cursor = restore_cursor.unwrap_or(true);
+
+        let result = self
+            .inner
+            .click_at_bounds(bounds, click_position, click_type.into(), restore_cursor)
+            .map(ClickResult::from)
+            .map_err(map_error);
+
+        // Get PID from process name if provided
+        let pid = process
+            .as_ref()
+            .and_then(|p| find_pid_for_process(&self.inner, p).ok());
+
+        // Capture screenshots if requested
+        let _screenshots = capture_screenshots(
+            &self.inner,
+            pid,
+            include_window_screenshot.unwrap_or(true) && pid.is_some(),
+            include_monitor_screenshots.unwrap_or(false),
+            "clickAtBounds",
+        );
+
+        result
+    }
+
+    /// Click on an element by its index from the last tree/vision query.
+    ///
+    /// This looks up cached bounds from the appropriate cache based on visionType,
+    /// then clicks at the specified position within those bounds.
+    ///
+    /// @param {number} index - 1-based index from the tree/vision output (e.g., #1, #2).
+    /// @param {VisionType} [visionType='UiTree'] - Source of the index: 'UiTree', 'Ocr', 'Omniparser', 'Gemini', or 'Dom'.
+    /// @param {number} [xPercentage=50] - X position within bounds as percentage (0-100).
+    /// @param {number} [yPercentage=50] - Y position within bounds as percentage (0-100).
+    /// @param {ClickType} [clickType='Left'] - Type of click: 'Left', 'Double', or 'Right'.
+    /// @param {boolean} [restoreCursor=true] - If true, restore cursor to original position after clicking.
+    /// @param {string} [process] - Process name for window screenshot capture. If provided, enables window screenshots.
+    /// @param {boolean} [includeWindowScreenshot=true] - Whether to capture window screenshot (requires process).
+    /// @param {boolean} [includeMonitorScreenshots=false] - Whether to capture monitor screenshots after clicking.
+    /// @returns {ClickResult} Result with clicked coordinates, element info, and method details.
+    #[napi]
+    #[allow(clippy::too_many_arguments)]
+    pub fn click_by_index(
+        &self,
+        index: u32,
+        vision_type: Option<VisionType>,
+        x_percentage: Option<u8>,
+        y_percentage: Option<u8>,
+        click_type: Option<ClickType>,
+        restore_cursor: Option<bool>,
+        process: Option<String>,
+        include_window_screenshot: Option<bool>,
+        include_monitor_screenshots: Option<bool>,
+    ) -> napi::Result<ClickResult> {
+        let vision_type = vision_type.unwrap_or(VisionType::UiTree);
+        let click_position = match (x_percentage, y_percentage) {
+            (Some(xp), Some(yp)) => Some((xp, yp)),
+            (Some(xp), None) => Some((xp, 50)),
+            (None, Some(yp)) => Some((50, yp)),
+            (None, None) => None,
+        };
+        let click_type = click_type.unwrap_or(ClickType::Left);
+        let restore_cursor = restore_cursor.unwrap_or(true);
+
+        let result = self
+            .inner
+            .click_by_index(
+                index,
+                vision_type.into(),
+                click_position,
+                click_type.into(),
+                restore_cursor,
+            )
+            .map(ClickResult::from)
+            .map_err(map_error);
+
+        // Get PID from process name if provided
+        let pid = process
+            .as_ref()
+            .and_then(|p| find_pid_for_process(&self.inner, p).ok());
+
+        // Capture screenshots if requested
+        let _screenshots = capture_screenshots(
+            &self.inner,
+            pid,
+            include_window_screenshot.unwrap_or(true) && pid.is_some(),
+            include_monitor_screenshots.unwrap_or(false),
+            "clickByIndex",
+        );
+
+        result
+    }
+
+    /// (async) Run a shell command.
+    ///
+    /// @param {string} [windowsCommand] - Command to run on Windows.
+    /// @param {string} [unixCommand] - Command to run on Unix.
+    /// @returns {Promise<CommandOutput>} The command output.
+    #[napi]
+    pub async fn run_command(
+        &self,
+        windows_command: Option<String>,
+        unix_command: Option<String>,
+    ) -> napi::Result<CommandOutput> {
+        self.inner
+            .run_command(windows_command.as_deref(), unix_command.as_deref())
+            .await
+            .map(|r| CommandOutput {
+                exit_status: r.exit_status,
+                stdout: r.stdout,
+                stderr: r.stderr,
+            })
+            .map_err(map_error)
+    }
+
+    /// (async) Execute a shell command using GitHub Actions-style syntax.
+    ///
+    /// @param {string} command - The command to run (can be single or multi-line).
+    /// @param {string} [shell] - Optional shell to use (defaults to PowerShell on Windows, bash on Unix).
+    /// @param {string} [workingDirectory] - Optional working directory for the command.
+    /// @returns {Promise<CommandOutput>} The command output.
+    #[napi]
+    pub async fn run(
+        &self,
+        command: String,
+        shell: Option<String>,
+        working_directory: Option<String>,
+    ) -> napi::Result<CommandOutput> {
+        self.inner
+            .run(
+                command.as_str(),
+                shell.as_deref(),
+                working_directory.as_deref(),
+            )
+            .await
+            .map(|r| CommandOutput {
+                exit_status: r.exit_status,
+                stdout: r.stdout,
+                stderr: r.stderr,
+            })
+            .map_err(map_error)
+    }
+
+    /// (async) Perform OCR on an image file.
+    ///
+    /// @param {string} imagePath - Path to the image file.
+    /// @returns {Promise<string>} The extracted text.
+    #[napi]
+    pub async fn ocr_image_path(&self, image_path: String) -> napi::Result<String> {
+        self.inner
+            .ocr_image_path(&image_path)
+            .await
+            .map_err(map_error)
+    }
+
+    /// (async) Perform OCR on a screenshot.
+    ///
+    /// @param {ScreenshotResult} screenshot - The screenshot to process.
+    /// @returns {Promise<string>} The extracted text.
+    #[napi]
+    pub async fn ocr_screenshot(&self, screenshot: ScreenshotResult) -> napi::Result<String> {
+        let rust_screenshot = screenshot.to_inner();
+        self.inner
+            .ocr_screenshot(&rust_screenshot)
+            .await
+            .map_err(map_error)
+    }
+
+    /// (async) Perform OCR on a window by process name and return structured results with bounding boxes.
+    /// Returns an OcrResult containing the OCR tree, formatted output, and index-to-bounds mapping
+    /// for click targeting.
+    ///
+    /// @param {string} process - Process name to match (e.g., 'chrome', 'notepad').
+    /// @param {boolean} [formatOutput=true] - Whether to generate formatted compact YAML output.
+    /// @returns {Promise<OcrResult>} Complete OCR result with tree, formatted output, and bounds mapping.
+    #[napi]
+    #[cfg(target_os = "windows")]
+    pub async fn perform_ocr_for_process(
+        &self,
+        process: String,
+        format_output: Option<bool>,
+    ) -> napi::Result<crate::types::OcrResult> {
+        let format_output = format_output.unwrap_or(true);
+
+        // Find PID for the process name
+        let pid = find_pid_for_process(&self.inner, &process)?;
+
+        // Find the application element by PID
+        let apps = self.inner.applications().map_err(map_error)?;
+        let window_element = apps
+            .into_iter()
+            .find(|app| app.process_id().ok() == Some(pid))
+            .ok_or_else(|| {
+                napi::Error::from_reason(format!("No window found for process '{}'", process))
+            })?;
+
+        // Get window bounds (absolute screen coordinates)
+        let bounds = window_element.bounds().map_err(map_error)?;
+        let (window_x, window_y, win_w, win_h) = bounds;
+
+        // Capture screenshot of the window
+        let screenshot = window_element.capture().map_err(map_error)?;
+
+        // Calculate DPI scale factors (physical screenshot pixels / logical window size)
+        let dpi_scale_w = screenshot.width as f64 / win_w;
+        let dpi_scale_h = screenshot.height as f64 / win_h;
+
+        // Perform OCR with bounding boxes
+        let ocr_element = self
+            .inner
+            .ocr_screenshot_with_bounds(&screenshot, window_x, window_y, dpi_scale_w, dpi_scale_h)
+            .map_err(map_error)?;
+
+        // Format the OCR tree if requested
+        let (formatted, index_to_bounds) = if format_output {
+            let result = computeruse::format_ocr_tree_as_compact_yaml(&ocr_element, 0);
+
+            // Populate the OCR cache for click_by_index support
+            self.inner
+                .populate_ocr_cache(result.index_to_bounds.clone());
+
+            let bounds_map: std::collections::HashMap<String, crate::types::OcrBoundsEntry> =
+                result
+                    .index_to_bounds
+                    .into_iter()
+                    .map(|(idx, (text, (x, y, w, h)))| {
+                        (
+                            idx.to_string(),
+                            crate::types::OcrBoundsEntry {
+                                text,
+                                bounds: crate::types::Bounds {
+                                    x,
+                                    y,
+                                    width: w,
+                                    height: h,
+                                },
+                            },
+                        )
+                    })
+                    .collect();
+            (Some(result.formatted), bounds_map)
+        } else {
+            (None, std::collections::HashMap::new())
+        };
+
+        let element_count = index_to_bounds.len() as u32;
+
+        Ok(crate::types::OcrResult {
+            tree: crate::types::OcrElement::from(ocr_element),
+            formatted,
+            index_to_bounds,
+            element_count,
+        })
+    }
+
+    /// (async) Perform OCR on a window by process name (non-Windows stub).
+    #[napi]
+    #[cfg(not(target_os = "windows"))]
+    pub async fn perform_ocr_for_process(
+        &self,
+        _process: String,
+        _format_output: Option<bool>,
+    ) -> napi::Result<crate::types::OcrResult> {
+        Err(napi::Error::from_reason(
+            "OCR with bounding boxes is currently only supported on Windows",
+        ))
+    }
+
+    /// (async) Capture DOM elements from the current browser tab.
+    ///
+    /// Extracts visible DOM elements with their properties and screen coordinates.
+    /// Uses JavaScript injection via Chrome extension to traverse the DOM tree.
+    ///
+    /// @param {number} [maxElements=200] - Maximum number of elements to capture.
+    /// @param {boolean} [formatOutput=true] - Whether to include formatted compact YAML output.
+    /// @returns {Promise<BrowserDomResult>} DOM elements with bounds for click targeting.
+    #[napi]
+    pub async fn capture_browser_dom(
+        &self,
+        max_elements: Option<u32>,
+        format_output: Option<bool>,
+    ) -> napi::Result<crate::types::BrowserDomResult> {
+        use std::collections::HashMap;
+        use std::time::Duration;
+
+        let max_elements = max_elements.unwrap_or(200);
+        let format_output = format_output.unwrap_or(true);
+
+        // Get viewport offset from Document element (more reliable than JS due to DPI scaling)
+        let viewport_offset = match self
+            .inner
+            .locator("role:Document")
+            .first(Some(Duration::from_millis(2000)))
+            .await
+        {
+            Ok(doc_element) => match doc_element.bounds() {
+                Ok((x, y, _w, _h)) => (x, y),
+                Err(_) => (0.0, 0.0),
+            },
+            Err(_) => (0.0, 0.0),
+        };
+
+        // JavaScript to extract visible DOM elements
+        let script = format!(
+            r#"
+(function() {{
+    const elements = [];
+    const maxElements = {max_elements};
+
+    const walker = document.createTreeWalker(
+        document.body,
+        NodeFilter.SHOW_ELEMENT,
+        {{
+            acceptNode: function(node) {{
+                const style = window.getComputedStyle(node);
+                const rect = node.getBoundingClientRect();
+
+                if (style.display === 'none' ||
+                    style.visibility === 'hidden' ||
+                    style.opacity === '0' ||
+                    rect.width === 0 ||
+                    rect.height === 0) {{
+                    return NodeFilter.FILTER_SKIP;
+                }}
+
+                return NodeFilter.FILTER_ACCEPT;
+            }}
+        }}
+    );
+
+    let node;
+    while (node = walker.nextNode()) {{
+        if (elements.length >= maxElements) {{
+            break;
+        }}
+
+        const rect = node.getBoundingClientRect();
+        const text = node.innerText ? node.innerText.substring(0, 100).trim() : null;
+
+        elements.push({{
+            tag: node.tagName.toLowerCase(),
+            id: node.id || null,
+            classes: Array.from(node.classList),
+            text: text,
+            href: node.href || null,
+            type: node.type || null,
+            name: node.name || null,
+            value: node.value || null,
+            placeholder: node.placeholder || null,
+            aria_label: node.getAttribute('aria-label'),
+            role: node.getAttribute('role'),
+            x: Math.round(rect.x * window.devicePixelRatio),
+            y: Math.round(rect.y * window.devicePixelRatio),
+            width: Math.round(rect.width * window.devicePixelRatio),
+            height: Math.round(rect.height * window.devicePixelRatio)
+        }});
+    }}
+
+    return JSON.stringify({{
+        elements: elements,
+        total_found: elements.length,
+        page_url: window.location.href,
+        page_title: document.title,
+        devicePixelRatio: window.devicePixelRatio
+    }});
+}})()"#
+        );
+
+        let result_str = self
+            .inner
+            .execute_browser_script(&script)
+            .await
+            .map_err(map_error)?;
+
+        let parsed: serde_json::Value = serde_json::from_str(&result_str)
+            .map_err(|e| napi::Error::from_reason(format!("Failed to parse DOM result: {e}")))?;
+
+        let page_url = parsed
+            .get("page_url")
+            .and_then(|v| v.as_str())
+            .unwrap_or("")
+            .to_string();
+        let page_title = parsed
+            .get("page_title")
+            .and_then(|v| v.as_str())
+            .unwrap_or("")
+            .to_string();
+
+        let raw_elements = parsed
+            .get("elements")
+            .and_then(|v| v.as_array())
+            .cloned()
+            .unwrap_or_default();
+
+        // Convert to BrowserDomElement and build index_to_bounds
+        let mut elements = Vec::new();
+        let mut index_to_bounds: HashMap<String, crate::types::DomBoundsEntry> = HashMap::new();
+        let mut formatted_lines: Vec<String> = Vec::new();
+
+        if format_output {
+            formatted_lines.push(format!(
+                "Browser DOM: {} elements (url: {}, title: {})",
+                raw_elements.len(),
+                page_url,
+                page_title
+            ));
+        }
+
+        for (i, elem) in raw_elements.iter().enumerate() {
+            let idx = i + 1;
+            let tag = elem
+                .get("tag")
+                .and_then(|v| v.as_str())
+                .unwrap_or("")
+                .to_string();
+            let id = elem.get("id").and_then(|v| v.as_str()).map(String::from);
+            let classes: Vec<String> = elem
+                .get("classes")
+                .and_then(|v| v.as_array())
+                .map(|arr| {
+                    arr.iter()
+                        .filter_map(|c| c.as_str().map(String::from))
+                        .collect()
+                })
+                .unwrap_or_default();
+            let text = elem.get("text").and_then(|v| v.as_str()).map(String::from);
+            let href = elem.get("href").and_then(|v| v.as_str()).map(String::from);
+            let r#type = elem.get("type").and_then(|v| v.as_str()).map(String::from);
+            let name = elem.get("name").and_then(|v| v.as_str()).map(String::from);
+            let value = elem.get("value").and_then(|v| v.as_str()).map(String::from);
+            let placeholder = elem
+                .get("placeholder")
+                .and_then(|v| v.as_str())
+                .map(String::from);
+            let aria_label = elem
+                .get("aria_label")
+                .and_then(|v| v.as_str())
+                .map(String::from);
+            let role = elem.get("role").and_then(|v| v.as_str()).map(String::from);
+
+            // Build bounds with viewport offset added
+            let x = elem.get("x").and_then(|v| v.as_f64()).unwrap_or(0.0) + viewport_offset.0;
+            let y = elem.get("y").and_then(|v| v.as_f64()).unwrap_or(0.0) + viewport_offset.1;
+            let width = elem.get("width").and_then(|v| v.as_f64()).unwrap_or(0.0);
+            let height = elem.get("height").and_then(|v| v.as_f64()).unwrap_or(0.0);
+
+            let bounds = crate::types::Bounds {
+                x,
+                y,
+                width,
+                height,
+            };
+
+            // Display name for index_to_bounds
+            let display_name = text
+                .as_ref()
+                .filter(|t| !t.is_empty())
+                .cloned()
+                .or_else(|| aria_label.clone())
+                .or_else(|| placeholder.clone())
+                .or_else(|| name.clone())
+                .or_else(|| id.clone())
+                .unwrap_or_else(|| format!("<{}>", tag));
+
+            // Format line for compact YAML
+            if format_output {
+                let mut line_parts = vec![format!("#{} [{}]", idx, tag.to_uppercase())];
+                if let Some(ref t) = text {
+                    if !t.is_empty() {
+                        let truncated = if t.len() > 40 {
+                            format!("{}...", &t[..40])
+                        } else {
+                            t.clone()
+                        };
+                        line_parts.push(truncated);
+                    }
+                }
+                if let Some(ref a) = aria_label {
+                    line_parts.push(format!("aria:{}", a));
+                }
+                if let Some(ref r) = role {
+                    line_parts.push(format!("role:{}", r));
+                }
+                formatted_lines.push(format!("  {}", line_parts.join(" ")));
+            }
+
+            index_to_bounds.insert(
+                idx.to_string(),
+                crate::types::DomBoundsEntry {
+                    name: display_name,
+                    tag: tag.clone(),
+                    bounds: bounds.clone(),
+                },
+            );
+
+            elements.push(crate::types::BrowserDomElement {
+                tag,
+                id,
+                classes,
+                text,
+                href,
+                r#type,
+                name,
+                value,
+                placeholder,
+                aria_label,
+                role,
+                bounds,
+            });
+        }
+
+        // Populate DOM cache for click_by_index
+        #[allow(clippy::type_complexity)]
+        let cache_items: std::collections::HashMap<
+            u32,
+            (String, String, (f64, f64, f64, f64)),
+        > = index_to_bounds
+            .iter()
+            .filter_map(|(key, entry)| {
+                key.parse::<u32>().ok().map(|idx| {
+                    (
+                        idx,
+                        (
+                            entry.name.clone(),
+                            entry.tag.clone(),
+                            (
+                                entry.bounds.x,
+                                entry.bounds.y,
+                                entry.bounds.width,
+                                entry.bounds.height,
+                            ),
+                        ),
+                    )
+                })
+            })
+            .collect();
+        self.inner.populate_dom_cache(cache_items);
+
+        Ok(crate::types::BrowserDomResult {
+            elements,
+            formatted: if format_output {
+                Some(formatted_lines.join("\n"))
+            } else {
+                None
+            },
+            index_to_bounds,
+            element_count: raw_elements.len() as u32,
+            page_url,
+            page_title,
+        })
+    }
+
+    /// (async) Get a clustered tree combining elements from multiple sources grouped by spatial proximity.
+    ///
+    /// Combines accessibility tree (UIA) elements with optional DOM, Omniparser, and Gemini Vision elements,
+    /// clustering nearby elements together. Each element is prefixed with its source:
+    /// - #u1, #u2... for UIA (accessibility tree)
+    /// - #d1, #d2... for DOM (browser content)
+    /// - #p1, #p2... for Omniparser (vision AI detection)
+    /// - #g1, #g2... for Gemini Vision (AI element detection)
+    ///
+    /// @param {string} process - Process name to match (e.g., 'chrome', 'notepad').
+    /// @param {number} [maxDomElements=100] - Maximum DOM elements to capture for browsers.
+    /// @param {boolean} [includeOmniparser=false] - Whether to include Omniparser vision detection.
+    /// @param {boolean} [includeGeminiVision=false] - Whether to include Gemini Vision AI detection.
+    /// @returns {Promise<ClusteredFormattingResult>} Clustered tree with prefixed indices.
+    #[napi]
+    pub async fn get_clustered_tree(
+        &self,
+        process: String,
+        max_dom_elements: Option<u32>,
+        include_omniparser: Option<bool>,
+        include_gemini_vision: Option<bool>,
+    ) -> napi::Result<crate::types::ClusteredFormattingResult> {
+        use std::collections::HashMap;
+
+        // Find PID for the process name
+        let pid = find_pid_for_process(&self.inner, &process)?;
+
+        let max_dom_elements = max_dom_elements.unwrap_or(100);
+        let include_omniparser = include_omniparser.unwrap_or(false);
+        let include_gemini_vision = include_gemini_vision.unwrap_or(false);
+
+        // Get UIA tree with bounds
+        let uia_result = self
+            .inner
+            .get_window_tree_result(pid, None, None)
+            .map_err(map_error)?;
+
+        // Build UIA bounds cache: HashMap<u32, (role, name, bounds, selector)>
+        #[allow(clippy::type_complexity)]
+        let mut uia_bounds: HashMap<
+            u32,
+            (String, String, (f64, f64, f64, f64), Option<String>),
+        > = HashMap::new();
+
+        // Use the formatted result to extract bounds
+        let formatted_result = computeruse::format_ui_node_as_compact_yaml(&uia_result.tree, 0);
+        for (idx, (role, name, bounds, selector)) in formatted_result.index_to_bounds {
+            uia_bounds.insert(idx, (role, name, bounds, selector));
+        }
+
+        // Check if this is a browser
+        let is_browser = computeruse::is_browser_process(pid);
+
+        // Build DOM bounds cache: HashMap<u32, (tag, identifier, bounds)>
+        #[allow(clippy::type_complexity)]
+        let mut dom_bounds: HashMap<u32, (String, String, (f64, f64, f64, f64))> = HashMap::new();
+
+        if is_browser {
+            // Try to capture DOM elements
+            match self
+                .capture_browser_dom(Some(max_dom_elements), Some(true))
+                .await
+            {
+                Ok(dom_result) => {
+                    for (idx_str, entry) in dom_result.index_to_bounds {
+                        if let Ok(idx) = idx_str.parse::<u32>() {
+                            let bounds = (
+                                entry.bounds.x,
+                                entry.bounds.y,
+                                entry.bounds.width,
+                                entry.bounds.height,
+                            );
+                            dom_bounds.insert(idx, (entry.tag, entry.name, bounds));
+                        }
+                    }
+                }
+                Err(_) => {
+                    // DOM capture failed (e.g., chrome:// page), continue with UIA only
+                }
+            }
+        }
+
+        // Build Omniparser items cache if requested
+        let mut omniparser_items: HashMap<u32, computeruse::OmniparserItem> = HashMap::new();
+
+        if include_omniparser {
+            match self
+                .perform_omniparser_for_process(process.clone(), None, Some(true))
+                .await
+            {
+                Ok(omni_result) => {
+                    for (idx_str, entry) in omni_result.index_to_bounds {
+                        if let Ok(idx) = idx_str.parse::<u32>() {
+                            omniparser_items.insert(
+                                idx,
+                                computeruse::OmniparserItem {
+                                    label: entry.label.clone(),
+                                    content: Some(entry.name.clone()),
+                                    box_2d: Some([
+                                        entry.bounds.x,
+                                        entry.bounds.y,
+                                        entry.bounds.x + entry.bounds.width,
+                                        entry.bounds.y + entry.bounds.height,
+                                    ]),
+                                },
+                            );
+                        }
+                    }
+                }
+                Err(_) => {
+                    // Omniparser failed, continue without it
+                }
+            }
+        }
+
+        // Build Gemini Vision items cache if requested
+        let mut vision_items: HashMap<u32, computeruse::VisionElement> = HashMap::new();
+
+        if include_gemini_vision {
+            match self
+                .perform_gemini_vision_for_process(process.clone(), Some(true))
+                .await
+            {
+                Ok(vision_result) => {
+                    for (idx_str, entry) in vision_result.index_to_bounds {
+                        if let Ok(idx) = idx_str.parse::<u32>() {
+                            vision_items.insert(
+                                idx,
+                                computeruse::VisionElement {
+                                    element_type: entry.element_type.clone(),
+                                    content: Some(entry.name.clone()),
+                                    description: None,
+                                    box_2d: Some([
+                                        entry.bounds.x,
+                                        entry.bounds.y,
+                                        entry.bounds.x + entry.bounds.width,
+                                        entry.bounds.y + entry.bounds.height,
+                                    ]),
+                                    interactivity: None,
+                                },
+                            );
+                        }
+                    }
+                }
+                Err(_) => {
+                    // Gemini Vision failed, continue without it
+                }
+            }
+        }
+
+        // Empty cache for OCR (not implemented yet)
+        #[allow(clippy::type_complexity)]
+        let ocr_bounds: HashMap<u32, (String, (f64, f64, f64, f64))> = HashMap::new();
+
+        // Call the core clustering function
+        let clustered_result = computeruse::format_clustered_tree_from_caches(
+            &uia_bounds,
+            &dom_bounds,
+            &ocr_bounds,
+            &omniparser_items,
+            &vision_items,
+        );
+
+        // Convert to SDK types
+        let mut index_to_source_and_bounds: HashMap<String, crate::types::ClusteredBoundsEntry> =
+            HashMap::new();
+
+        for (key, (source, original_idx, (x, y, w, h))) in
+            clustered_result.index_to_source_and_bounds
+        {
+            let sdk_source = match source {
+                computeruse::ElementSource::Uia => crate::types::ElementSource::Uia,
+                computeruse::ElementSource::Dom => crate::types::ElementSource::Dom,
+                computeruse::ElementSource::Ocr => crate::types::ElementSource::Ocr,
+                computeruse::ElementSource::Omniparser => crate::types::ElementSource::Omniparser,
+                computeruse::ElementSource::Gemini => crate::types::ElementSource::Gemini,
+            };
+            index_to_source_and_bounds.insert(
+                key,
+                crate::types::ClusteredBoundsEntry {
+                    source: sdk_source,
+                    original_index: original_idx,
+                    bounds: crate::types::Bounds {
+                        x,
+                        y,
+                        width: w,
+                        height: h,
+                    },
+                },
+            );
+        }
+
+        Ok(crate::types::ClusteredFormattingResult {
+            formatted: clustered_result.formatted,
+            index_to_source_and_bounds,
+        })
+    }
+
+    /// (async) Perform Gemini vision AI detection on a window by process name.
+    ///
+    /// Captures a screenshot and sends it to the Gemini vision backend for UI element detection.
+    /// Requires GEMINI_VISION_BACKEND_URL environment variable (defaults to https://app.mediar.ai/api/vision/parse).
+    ///
+    /// @param {string} process - Process name to match (e.g., 'chrome', 'notepad').
+    /// @param {boolean} [formatOutput=true] - Whether to include formatted compact YAML output.
+    /// @returns {Promise<GeminiVisionResult>} Detected UI elements with bounds for click targeting.
+    #[napi]
+    pub async fn perform_gemini_vision_for_process(
+        &self,
+        process: String,
+        format_output: Option<bool>,
+    ) -> napi::Result<crate::types::GeminiVisionResult> {
+        use base64::{engine::general_purpose, Engine};
+        use image::imageops::FilterType;
+        use image::{codecs::png::PngEncoder, ExtendedColorType, ImageBuffer, ImageEncoder, Rgba};
+        use std::collections::HashMap;
+        use std::io::Cursor;
+
+        let format_output = format_output.unwrap_or(true);
+
+        // Find PID for the process name
+        let pid = find_pid_for_process(&self.inner, &process)?;
+
+        // Find the window element for this process
+        let apps = self.inner.applications().map_err(map_error)?;
+        let window_element = apps
+            .into_iter()
+            .find(|app| app.process_id().ok() == Some(pid))
+            .ok_or_else(|| {
+                napi::Error::from_reason(format!("No window found for process '{}'", process))
+            })?;
+
+        // Get window bounds
+        let bounds = window_element.bounds().map_err(map_error)?;
+        let (window_x, window_y, win_w, win_h) = bounds;
+
+        // Capture screenshot
+        let screenshot = window_element.capture().map_err(map_error)?;
+        let original_width = screenshot.width;
+        let original_height = screenshot.height;
+
+        // Calculate DPI scale
+        let dpi_scale_w = original_width as f64 / win_w;
+        let dpi_scale_h = original_height as f64 / win_h;
+
+        // Convert BGRA to RGBA
+        let rgba_data: Vec<u8> = screenshot
+            .image_data
+            .chunks_exact(4)
+            .flat_map(|bgra| [bgra[2], bgra[1], bgra[0], bgra[3]])
+            .collect();
+
+        // Resize if needed (max 1920px)
+        const MAX_DIM: u32 = 1920;
+        let (final_width, final_height, final_rgba_data, scale_factor) = if original_width > MAX_DIM
+            || original_height > MAX_DIM
+        {
+            let scale = (MAX_DIM as f32 / original_width.max(original_height) as f32).min(1.0);
+            let new_width = (original_width as f32 * scale).round() as u32;
+            let new_height = (original_height as f32 * scale).round() as u32;
+
+            let img =
+                ImageBuffer::<Rgba<u8>, _>::from_raw(original_width, original_height, rgba_data)
+                    .ok_or_else(|| napi::Error::from_reason("Failed to create image buffer"))?;
+
+            let resized =
+                image::imageops::resize(&img, new_width, new_height, FilterType::Lanczos3);
+            (new_width, new_height, resized.into_raw(), scale as f64)
+        } else {
+            (original_width, original_height, rgba_data, 1.0)
+        };
+
+        // Encode to PNG
+        let mut png_data = Vec::new();
+        let encoder = PngEncoder::new(Cursor::new(&mut png_data));
+        encoder
+            .write_image(
+                &final_rgba_data,
+                final_width,
+                final_height,
+                ExtendedColorType::Rgba8,
+            )
+            .map_err(|e| napi::Error::from_reason(format!("Failed to encode PNG: {e}")))?;
+
+        let base64_image = general_purpose::STANDARD.encode(&png_data);
+
+        // Call Gemini Vision backend
+        let backend_url = std::env::var("GEMINI_VISION_BACKEND_URL")
+            .unwrap_or_else(|_| "https://app.mediar.ai/api/vision/parse".to_string());
+
+        let client = reqwest::Client::builder()
+            .timeout(std::time::Duration::from_secs(300))
+            .build()
+            .map_err(|e| napi::Error::from_reason(format!("Failed to create HTTP client: {e}")))?;
+
+        let payload = serde_json::json!({
+            "image": base64_image,
+            "model": "gemini",
+            "prompt": "Detect all UI elements in this screenshot. Return their type, content, description, bounding boxes, and interactivity."
+        });
+
+        let resp = client
+            .post(&backend_url)
+            .header("Content-Type", "application/json")
+            .json(&payload)
+            .send()
+            .await
+            .map_err(|e| napi::Error::from_reason(format!("Vision backend request failed: {e}")))?;
+
+        if !resp.status().is_success() {
+            let text = resp.text().await.unwrap_or_default();
+            return Err(napi::Error::from_reason(format!(
+                "Vision backend error: {}",
+                text
+            )));
+        }
+
+        let response_text = resp
+            .text()
+            .await
+            .map_err(|e| napi::Error::from_reason(format!("Failed to read response: {e}")))?;
+
+        let parsed: serde_json::Value = serde_json::from_str(&response_text)
+            .map_err(|e| napi::Error::from_reason(format!("Failed to parse response: {e}")))?;
+
+        if let Some(error) = parsed.get("error").and_then(|v| v.as_str()) {
+            return Err(napi::Error::from_reason(format!("Vision error: {}", error)));
+        }
+
+        let raw_elements = parsed
+            .get("elements")
+            .and_then(|v| v.as_array())
+            .cloned()
+            .unwrap_or_default();
+
+        // Convert to VisionElement with absolute screen coordinates
+        let mut elements = Vec::new();
+        let mut index_to_bounds: HashMap<String, crate::types::VisionBoundsEntry> = HashMap::new();
+        let mut formatted_lines: Vec<String> = Vec::new();
+
+        if format_output {
+            formatted_lines.push(format!(
+                "Gemini Vision: {} elements (PID: {})",
+                raw_elements.len(),
+                pid
+            ));
+        }
+
+        let inv_scale = 1.0 / scale_factor;
+
+        for (i, elem) in raw_elements.iter().enumerate() {
+            let idx = i + 1;
+            let element_type = elem
+                .get("type")
+                .and_then(|v| v.as_str())
+                .unwrap_or("unknown")
+                .to_string();
+            let content = elem
+                .get("content")
+                .and_then(|v| v.as_str())
+                .filter(|s| !s.is_empty())
+                .map(String::from);
+            let description = elem
+                .get("description")
+                .and_then(|v| v.as_str())
+                .filter(|s| !s.is_empty())
+                .map(String::from);
+            let interactivity = elem.get("interactivity").and_then(|v| v.as_bool());
+
+            // Get normalized bbox [x1, y1, x2, y2] from 0-1
+            let bbox = elem.get("bbox").and_then(|v| v.as_array());
+            let bounds = bbox.and_then(|arr| {
+                if arr.len() >= 4 {
+                    let x1 = arr[0].as_f64()? * final_width as f64;
+                    let y1 = arr[1].as_f64()? * final_height as f64;
+                    let x2 = arr[2].as_f64()? * final_width as f64;
+                    let y2 = arr[3].as_f64()? * final_height as f64;
+
+                    // Scale back to original size and convert to logical screen coords
+                    let abs_x = window_x + (x1 * inv_scale / dpi_scale_w);
+                    let abs_y = window_y + (y1 * inv_scale / dpi_scale_h);
+                    let abs_w = (x2 - x1) * inv_scale / dpi_scale_w;
+                    let abs_h = (y2 - y1) * inv_scale / dpi_scale_h;
+
+                    Some(crate::types::Bounds {
+                        x: abs_x,
+                        y: abs_y,
+                        width: abs_w,
+                        height: abs_h,
+                    })
+                } else {
+                    None
+                }
+            });
+
+            // Display name for index_to_bounds
+            let display_name = content
+                .as_ref()
+                .cloned()
+                .or_else(|| description.clone())
+                .unwrap_or_else(|| format!("<{}>", element_type));
+
+            // Format line for compact YAML
+            if format_output {
+                let mut line_parts = vec![format!("#{} [{}]", idx, element_type.to_uppercase())];
+                if let Some(ref c) = content {
+                    let truncated = if c.len() > 40 {
+                        format!("{}...", &c[..40])
+                    } else {
+                        c.clone()
+                    };
+                    line_parts.push(truncated);
+                }
+                if let Some(ref d) = description {
+                    let truncated = if d.len() > 30 {
+                        format!("{}...", &d[..30])
+                    } else {
+                        d.clone()
+                    };
+                    line_parts.push(format!("desc:{}", truncated));
+                }
+                if interactivity == Some(true) {
+                    line_parts.push("interactive".to_string());
+                }
+                formatted_lines.push(format!("  {}", line_parts.join(" ")));
+            }
+
+            if let Some(ref b) = bounds {
+                index_to_bounds.insert(
+                    idx.to_string(),
+                    crate::types::VisionBoundsEntry {
+                        name: display_name.clone(),
+                        element_type: element_type.clone(),
+                        bounds: b.clone(),
+                    },
+                );
+            }
+
+            elements.push(crate::types::VisionElement {
+                element_type,
+                content,
+                description,
+                bounds,
+                interactivity,
+            });
+        }
+
+        // Populate the Vision cache for click_by_index support
+        let cache_items: HashMap<u32, computeruse::VisionElement> = elements
+            .iter()
+            .enumerate()
+            .map(|(i, elem)| {
+                let box_2d = elem
+                    .bounds
+                    .as_ref()
+                    .map(|b| [b.x, b.y, b.x + b.width, b.y + b.height]);
+                (
+                    (i + 1) as u32,
+                    computeruse::VisionElement {
+                        element_type: elem.element_type.clone(),
+                        content: elem.content.clone(),
+                        description: elem.description.clone(),
+                        box_2d,
+                        interactivity: elem.interactivity,
+                    },
+                )
+            })
+            .collect();
+        self.inner.populate_vision_cache(cache_items);
+
+        Ok(crate::types::GeminiVisionResult {
+            elements,
+            formatted: if format_output {
+                Some(formatted_lines.join("\n"))
+            } else {
+                None
+            },
+            index_to_bounds,
+            element_count: raw_elements.len() as u32,
+        })
+    }
+
+    /// (async) Perform Omniparser V2 detection on a window by process name.
+    ///
+    /// Captures a screenshot and sends it to the Omniparser backend for icon/field detection.
+    /// Requires OMNIPARSER_BACKEND_URL environment variable (defaults to https://app.mediar.ai/api/omniparser/parse).
+    ///
+    /// @param {string} process - Process name to match (e.g., 'chrome', 'notepad').
+    /// @param {number} [imgsz=1920] - Icon detection image size (640-1920). Higher = better but slower.
+    /// @param {boolean} [formatOutput=true] - Whether to include formatted compact YAML output.
+    /// @returns {Promise<OmniparserResult>} Detected items with bounds for click targeting.
+    #[napi]
+    pub async fn perform_omniparser_for_process(
+        &self,
+        process: String,
+        imgsz: Option<u32>,
+        format_output: Option<bool>,
+    ) -> napi::Result<crate::types::OmniparserResult> {
+        use base64::{engine::general_purpose, Engine};
+        use image::imageops::FilterType;
+        use image::{codecs::png::PngEncoder, ExtendedColorType, ImageBuffer, ImageEncoder, Rgba};
+        use std::collections::HashMap;
+        use std::io::Cursor;
+
+        let imgsz = imgsz.unwrap_or(1920).clamp(640, 1920);
+        let format_output = format_output.unwrap_or(true);
+
+        // Find PID for the process name
+        let pid = find_pid_for_process(&self.inner, &process)?;
+
+        // Find the window element for this process
+        let apps = self.inner.applications().map_err(map_error)?;
+        let window_element = apps
+            .into_iter()
+            .find(|app| app.process_id().ok() == Some(pid))
+            .ok_or_else(|| {
+                napi::Error::from_reason(format!("No window found for process '{}'", process))
+            })?;
+
+        // Get window bounds
+        let bounds = window_element.bounds().map_err(map_error)?;
+        let (window_x, window_y, win_w, win_h) = bounds;
+
+        // Capture screenshot
+        let screenshot = window_element.capture().map_err(map_error)?;
+        let original_width = screenshot.width;
+        let original_height = screenshot.height;
+
+        // Calculate DPI scale
+        let dpi_scale_w = original_width as f64 / win_w;
+        let dpi_scale_h = original_height as f64 / win_h;
+
+        // Convert BGRA to RGBA
+        let rgba_data: Vec<u8> = screenshot
+            .image_data
+            .chunks_exact(4)
+            .flat_map(|bgra| [bgra[2], bgra[1], bgra[0], bgra[3]])
+            .collect();
+
+        // Resize if needed (max 1920px)
+        const MAX_DIM: u32 = 1920;
+        let (final_width, final_height, final_rgba_data, scale_factor) = if original_width > MAX_DIM
+            || original_height > MAX_DIM
+        {
+            let scale = (MAX_DIM as f32 / original_width.max(original_height) as f32).min(1.0);
+            let new_width = (original_width as f32 * scale).round() as u32;
+            let new_height = (original_height as f32 * scale).round() as u32;
+
+            let img =
+                ImageBuffer::<Rgba<u8>, _>::from_raw(original_width, original_height, rgba_data)
+                    .ok_or_else(|| napi::Error::from_reason("Failed to create image buffer"))?;
+
+            let resized =
+                image::imageops::resize(&img, new_width, new_height, FilterType::Lanczos3);
+            (new_width, new_height, resized.into_raw(), scale as f64)
+        } else {
+            (original_width, original_height, rgba_data, 1.0)
+        };
+
+        // Encode to PNG
+        let mut png_data = Vec::new();
+        let encoder = PngEncoder::new(Cursor::new(&mut png_data));
+        encoder
+            .write_image(
+                &final_rgba_data,
+                final_width,
+                final_height,
+                ExtendedColorType::Rgba8,
+            )
+            .map_err(|e| napi::Error::from_reason(format!("Failed to encode PNG: {e}")))?;
+
+        let base64_image = general_purpose::STANDARD.encode(&png_data);
+
+        // Call Omniparser backend
+        let backend_url = std::env::var("OMNIPARSER_BACKEND_URL")
+            .unwrap_or_else(|_| "https://app.mediar.ai/api/omniparser/parse".to_string());
+
+        let client = reqwest::Client::builder()
+            .timeout(std::time::Duration::from_secs(300))
+            .build()
+            .map_err(|e| napi::Error::from_reason(format!("Failed to create HTTP client: {e}")))?;
+
+        let payload = serde_json::json!({
+            "image": base64_image,
+            "imgsz": imgsz
+        });
+
+        let resp = client
+            .post(&backend_url)
+            .header("Content-Type", "application/json")
+            .json(&payload)
+            .send()
+            .await
+            .map_err(|e| {
+                napi::Error::from_reason(format!("Omniparser backend request failed: {e}"))
+            })?;
+
+        if !resp.status().is_success() {
+            let text = resp.text().await.unwrap_or_default();
+            return Err(napi::Error::from_reason(format!(
+                "Omniparser backend error: {}",
+                text
+            )));
+        }
+
+        let response_text = resp
+            .text()
+            .await
+            .map_err(|e| napi::Error::from_reason(format!("Failed to read response: {e}")))?;
+
+        let parsed: serde_json::Value = serde_json::from_str(&response_text)
+            .map_err(|e| napi::Error::from_reason(format!("Failed to parse response: {e}")))?;
+
+        if let Some(error) = parsed.get("error").and_then(|v| v.as_str()) {
+            return Err(napi::Error::from_reason(format!(
+                "Omniparser error: {}",
+                error
+            )));
+        }
+
+        let raw_elements = parsed
+            .get("elements")
+            .and_then(|v| v.as_array())
+            .cloned()
+            .unwrap_or_default();
+
+        // Convert to OmniparserItem with absolute screen coordinates
+        let mut items = Vec::new();
+        let mut index_to_bounds: HashMap<String, crate::types::OmniparserBoundsEntry> =
+            HashMap::new();
+        let mut formatted_lines: Vec<String> = Vec::new();
+
+        if format_output {
+            formatted_lines.push(format!(
+                "Omniparser: {} items (PID: {})",
+                raw_elements.len(),
+                pid
+            ));
+        }
+
+        let inv_scale = 1.0 / scale_factor;
+
+        for (i, elem) in raw_elements.iter().enumerate() {
+            let idx = i + 1;
+            let label = elem
+                .get("type")
+                .and_then(|v| v.as_str())
+                .unwrap_or("unknown")
+                .to_string();
+            let content = elem
+                .get("content")
+                .and_then(|v| v.as_str())
+                .filter(|s| !s.is_empty())
+                .map(String::from);
+
+            // Get normalized bbox [x1, y1, x2, y2] from 0-1
+            let bbox = elem.get("bbox").and_then(|v| v.as_array());
+            let bounds = bbox.and_then(|arr| {
+                if arr.len() >= 4 {
+                    let x1 = arr[0].as_f64()? * final_width as f64;
+                    let y1 = arr[1].as_f64()? * final_height as f64;
+                    let x2 = arr[2].as_f64()? * final_width as f64;
+                    let y2 = arr[3].as_f64()? * final_height as f64;
+
+                    // Scale back to original size and convert to logical screen coords
+                    let abs_x = window_x + (x1 * inv_scale / dpi_scale_w);
+                    let abs_y = window_y + (y1 * inv_scale / dpi_scale_h);
+                    let abs_w = (x2 - x1) * inv_scale / dpi_scale_w;
+                    let abs_h = (y2 - y1) * inv_scale / dpi_scale_h;
+
+                    Some(crate::types::Bounds {
+                        x: abs_x,
+                        y: abs_y,
+                        width: abs_w,
+                        height: abs_h,
+                    })
+                } else {
+                    None
+                }
+            });
+
+            // Display name for index_to_bounds
+            let display_name = content
+                .as_ref()
+                .cloned()
+                .unwrap_or_else(|| format!("<{}>", label));
+
+            // Format line for compact YAML
+            if format_output {
+                let mut line_parts = vec![format!("#{} [{}]", idx, label.to_uppercase())];
+                if let Some(ref c) = content {
+                    let truncated = if c.len() > 50 {
+                        format!("{}...", &c[..50])
+                    } else {
+                        c.clone()
+                    };
+                    line_parts.push(truncated);
+                }
+                formatted_lines.push(format!("  {}", line_parts.join(" ")));
+            }
+
+            if let Some(ref b) = bounds {
+                index_to_bounds.insert(
+                    idx.to_string(),
+                    crate::types::OmniparserBoundsEntry {
+                        name: display_name.clone(),
+                        label: label.clone(),
+                        bounds: b.clone(),
+                    },
+                );
+            }
+
+            items.push(crate::types::OmniparserItem {
+                label,
+                content,
+                bounds,
+            });
+        }
+
+        // Populate the Omniparser cache for click_by_index support
+        let cache_items: HashMap<u32, computeruse::OmniparserItem> = items
+            .iter()
+            .enumerate()
+            .map(|(i, item)| {
+                let box_2d = item
+                    .bounds
+                    .as_ref()
+                    .map(|b| [b.x, b.y, b.x + b.width, b.y + b.height]);
+                (
+                    (i + 1) as u32,
+                    computeruse::OmniparserItem {
+                        label: item.label.clone(),
+                        content: item.content.clone(),
+                        box_2d,
+                    },
+                )
+            })
+            .collect();
+        self.inner.populate_omniparser_cache(cache_items);
+
+        Ok(crate::types::OmniparserResult {
+            items,
+            formatted: if format_output {
+                Some(formatted_lines.join("\n"))
+            } else {
+                None
+            },
+            index_to_bounds,
+            item_count: raw_elements.len() as u32,
+        })
+    }
+
+    /// (async) Get the currently focused browser window.
+    ///
+    /// @returns {Promise<Element>} The current browser window element.
+    #[napi]
+    pub async fn get_current_browser_window(&self) -> napi::Result<Element> {
+        self.inner
+            .get_current_browser_window()
+            .await
+            .map(Element::from)
+            .map_err(map_error)
+    }
+
+    /// Create a locator for finding UI elements.
+    ///
+    /// @param {string | Selector} selector - The selector.
+    /// @returns {Locator} A locator for finding elements.
+    #[napi]
+    pub fn locator(
+        &self,
+        #[napi(ts_arg_type = "string | Selector")] selector: Either<String, &Selector>,
+    ) -> napi::Result<Locator> {
+        use napi::bindgen_prelude::Either::*;
+        let sel_rust: computeruse::selector::Selector = match selector {
+            A(sel_str) => sel_str.as_str().into(),
+            B(sel_obj) => sel_obj.inner.clone(),
+        };
+        let loc = self.inner.locator(sel_rust);
+        Ok(Locator::from(loc))
+    }
+
+    /// Create a process-scoped locator for finding UI elements.
+    /// This is the recommended way to create locators - always scope to a specific process.
+    ///
+    /// @param {string} process - Process name to scope the search (e.g., 'chrome', 'notepad').
+    /// @param {string | Selector} selector - The selector to find within the process.
+    /// @param {string} [windowSelector] - Optional window selector for additional filtering.
+    /// @returns {Locator} A locator for finding elements within the process.
+    #[napi]
+    pub fn locator_for_process(
+        &self,
+        process: String,
+        #[napi(ts_arg_type = "string | Selector")] selector: Either<String, &Selector>,
+        window_selector: Option<String>,
+    ) -> napi::Result<Locator> {
+        use napi::bindgen_prelude::Either::*;
+
+        // Build the full selector string like MCP does
+        let selector_str = match &selector {
+            A(sel_str) => sel_str.clone(),
+            B(sel_obj) => format!("{:?}", sel_obj.inner),
+        };
+
+        let full_selector = if selector_str.is_empty() {
+            if let Some(window_sel) = window_selector {
+                format!("process:{} >> {}", process, window_sel)
+            } else {
+                format!("process:{}", process)
+            }
+        } else if let Some(window_sel) = window_selector {
+            format!("process:{} >> {} >> {}", process, window_sel, selector_str)
+        } else {
+            format!("process:{} >> {}", process, selector_str)
+        };
+
+        let sel_rust: computeruse::selector::Selector = full_selector.as_str().into();
+        let loc = self.inner.locator(sel_rust);
+        Ok(Locator::from(loc))
+    }
+
+    /// (async) Get the currently focused window.
+    ///
+    /// @returns {Promise<Element>} The current window element.
+    #[napi]
+    pub async fn get_current_window(&self) -> napi::Result<Element> {
+        self.inner
+            .get_current_window()
+            .await
+            .map(Element::from)
+            .map_err(map_error)
+    }
+
+    /// (async) Get the currently focused application.
+    ///
+    /// @returns {Promise<Element>} The current application element.
+    #[napi]
+    pub async fn get_current_application(&self) -> napi::Result<Element> {
+        self.inner
+            .get_current_application()
+            .await
+            .map(Element::from)
+            .map_err(map_error)
+    }
+
+    /// Get the currently focused element.
+    ///
+    /// @returns {Element} The focused element.
+    #[napi]
+    pub fn focused_element(&self) -> napi::Result<Element> {
+        self.inner
+            .focused_element()
+            .map(Element::from)
+            .map_err(map_error)
+    }
+
+    /// Open a URL in a browser.
+    ///
+    /// @param {string} url - The URL to open.
+    /// @param {string} [browser] - The browser to use. Can be "Default", "Chrome", "Firefox", "Edge", "Brave", "Opera", "Vivaldi", or a custom browser path.
+    /// @param {boolean} [includeWindowScreenshot=true] - Whether to capture window screenshot after opening
+    /// @param {boolean} [includeMonitorScreenshots=false] - Whether to capture monitor screenshots after opening
+    #[napi]
+    pub fn open_url(
+        &self,
+        url: String,
+        browser: Option<String>,
+        include_window_screenshot: Option<bool>,
+        include_monitor_screenshots: Option<bool>,
+    ) -> napi::Result<Element> {
+        let browser_enum = browser.map(|b| match b.to_lowercase().as_str() {
+            "default" => computeruse::Browser::Default,
+            "chrome" => computeruse::Browser::Chrome,
+            "firefox" => computeruse::Browser::Firefox,
+            "edge" => computeruse::Browser::Edge,
+            "brave" => computeruse::Browser::Brave,
+            "opera" => computeruse::Browser::Opera,
+            "vivaldi" => computeruse::Browser::Vivaldi,
+            custom => computeruse::Browser::Custom(custom.to_string()),
+        });
+        let element = self.inner.open_url(&url, browser_enum).map_err(map_error)?;
+
+        // Capture screenshots if enabled (window default: true, monitor default: false)
+        let _screenshots = capture_screenshots(
+            &self.inner,
+            element.process_id().ok(),
+            include_window_screenshot.unwrap_or(true),
+            include_monitor_screenshots.unwrap_or(false),
+            "openUrl",
+        );
+
+        Ok(Element::from(element))
+    }
+
+    /// Open a file with its default application.
+    ///
+    /// @param {string} filePath - Path to the file to open.
+    /// @param {string} [process] - Process name for window screenshot capture. If provided, enables window screenshots.
+    /// @param {boolean} [includeWindowScreenshot=true] - Whether to capture window screenshot (requires process).
+    /// @param {boolean} [includeMonitorScreenshots=false] - Whether to capture monitor screenshots after opening.
+    #[napi]
+    pub fn open_file(
+        &self,
+        file_path: String,
+        process: Option<String>,
+        include_window_screenshot: Option<bool>,
+        include_monitor_screenshots: Option<bool>,
+    ) -> napi::Result<()> {
+        let result = self.inner.open_file(&file_path).map_err(map_error);
+
+        // Get PID from process name if provided
+        let pid = process
+            .as_ref()
+            .and_then(|p| find_pid_for_process(&self.inner, p).ok());
+
+        // Capture screenshots if requested
+        let _screenshots = capture_screenshots(
+            &self.inner,
+            pid,
+            include_window_screenshot.unwrap_or(true) && pid.is_some(),
+            include_monitor_screenshots.unwrap_or(false),
+            "openFile",
+        );
+
+        result
+    }
+
+    /// Activate a browser window by title.
+    ///
+    /// @param {string} title - The window title to match.
+    #[napi]
+    pub fn activate_browser_window_by_title(&self, title: String) -> napi::Result<()> {
+        self.inner
+            .activate_browser_window_by_title(&title)
+            .map_err(map_error)
+    }
+
+    /// Get the UI tree for a window identified by process name and optional title.
+    ///
+    /// @param {string} process - Process name to match (e.g., 'chrome', 'notepad').
+    /// @param {string} [title] - Optional window title filter.
+    /// @param {TreeBuildConfig} [config] - Optional configuration for tree building.
+    /// @returns {UINode} Complete UI tree starting from the identified window.
+    #[napi]
+    pub fn get_window_tree(
+        &self,
+        process: String,
+        title: Option<String>,
+        config: Option<TreeBuildConfig>,
+    ) -> napi::Result<UINode> {
+        // Find PID for the process name
+        let pid = find_pid_for_process(&self.inner, &process)?;
+
+        let rust_config = config.map(|c| c.into());
+        self.inner
+            .get_window_tree(pid, title.as_deref(), rust_config)
+            .map(UINode::from)
+            .map_err(map_error)
+    }
+
+    /// Get the UI tree with full result including formatting and bounds mapping.
+    ///
+    /// This is the recommended method for getting window trees when you need:
+    /// - Formatted YAML output for LLM consumption
+    /// - Index-to-bounds mapping for click targeting
+    /// - Browser detection
+    ///
+    /// @param {string} process - Process name to match (e.g., 'chrome', 'notepad').
+    /// @param {string} [title] - Optional window title filter.
+    /// @param {TreeBuildConfig} [config] - Configuration options:
+    ///   - formatOutput: Enable formatted output (default: true if treeOutputFormat set)
+    ///   - treeOutputFormat: 'CompactYaml' (default) or 'VerboseJson'
+    ///   - treeFromSelector: Selector to start tree from (use getWindowTreeResultAsync for this)
+    ///   - includeWindowScreenshot: Save window screenshot to executions dir (default: false)
+    ///   - includeMonitorScreenshots: Save all monitor screenshots to executions dir (default: false)
+    /// @returns {WindowTreeResult} Complete result with tree, formatted output, bounds mapping, and screenshot paths.
+    #[napi]
+    pub fn get_window_tree_result(
+        &self,
+        process: String,
+        title: Option<String>,
+        config: Option<TreeBuildConfig>,
+    ) -> napi::Result<WindowTreeResult> {
+        // Find PID for the process name
+        let pid = find_pid_for_process(&self.inner, &process)?;
+
+        // Extract screenshot options (window: true, monitor: false by default)
+        let include_window_screenshot = config
+            .as_ref()
+            .and_then(|c| c.include_window_screenshot)
+            .unwrap_or(true);
+        let include_monitor_screenshots = config
+            .as_ref()
+            .and_then(|c| c.include_monitor_screenshots)
+            .unwrap_or(false);
+
+        // Extract options before converting config
+        let output_format = config
+            .as_ref()
+            .and_then(|c| c.tree_output_format)
+            .unwrap_or(TreeOutputFormat::CompactYaml);
+
+        // If format is VerboseJson, we don't need formatted output from core
+        // ClusteredYaml is treated like CompactYaml (needs format_output = true)
+        let rust_config = config.map(|mut c| {
+            if matches!(output_format, TreeOutputFormat::VerboseJson) {
+                c.format_output = Some(false);
+            } else if c.format_output.is_none() {
+                c.format_output = Some(true);
+            }
+            c.into()
+        });
+
+        let result = self
+            .inner
+            .get_window_tree_result(pid, title.as_deref(), rust_config)
+            .map_err(map_error)?;
+
+        // Convert and handle format
+        let mut sdk_result = WindowTreeResult::from(result);
+
+        // For VerboseJson, serialize the tree as the formatted output
+        if matches!(output_format, TreeOutputFormat::VerboseJson) {
+            sdk_result.formatted =
+                Some(serde_json::to_string_pretty(&sdk_result.tree).unwrap_or_default());
+        }
+
+        // Handle screenshot capture and saving using helper
+        let screenshots = capture_screenshots(
+            &self.inner,
+            Some(pid),
+            include_window_screenshot,
+            include_monitor_screenshots,
+            "getWindowTreeResult",
+        );
+        sdk_result.window_screenshot_path = screenshots.window_path;
+        sdk_result.monitor_screenshot_paths = screenshots.monitor_paths;
+
+        Ok(sdk_result)
+    }
+
+    /// (async) Get the UI tree with full result, supporting tree_from_selector.
+    ///
+    /// Use this method when you need to scope the tree to a specific subtree using a selector.
+    ///
+    /// @param {string} process - Process name to match (e.g., 'chrome', 'notepad').
+    /// @param {string} [title] - Optional window title filter.
+    /// @param {TreeBuildConfig} [config] - Configuration options:
+    ///   - formatOutput: Enable formatted output (default: true)
+    ///   - treeOutputFormat: 'CompactYaml' (default) or 'VerboseJson'
+    ///   - treeFromSelector: Selector to start tree from (e.g., "role:Dialog")
+    /// @returns {Promise<WindowTreeResult>} Complete result with tree, formatted output, and bounds mapping.
+    #[napi]
+    pub async fn get_window_tree_result_async(
+        &self,
+        process: String,
+        title: Option<String>,
+        config: Option<TreeBuildConfig>,
+    ) -> napi::Result<WindowTreeResult> {
+        use std::collections::HashMap;
+
+        // Find PID for the process name
+        let pid = find_pid_for_process(&self.inner, &process)?;
+
+        // Extract vision and format options from config
+        let include_gemini_vision = config
+            .as_ref()
+            .and_then(|c| c.include_gemini_vision)
+            .unwrap_or(false);
+        let include_omniparser = config
+            .as_ref()
+            .and_then(|c| c.include_omniparser)
+            .unwrap_or(false);
+        let include_ocr = config.as_ref().and_then(|c| c.include_ocr).unwrap_or(false);
+        let include_browser_dom = config
+            .as_ref()
+            .and_then(|c| c.include_browser_dom)
+            .unwrap_or(false);
+        let output_format = config
+            .as_ref()
+            .and_then(|c| c.tree_output_format)
+            .unwrap_or(TreeOutputFormat::CompactYaml);
+
+        let has_vision_options =
+            include_gemini_vision || include_omniparser || include_ocr || include_browser_dom;
+
+        // Build rust config with from_selector passed through
+        let rust_config = config.as_ref().map(|c| {
+            let mut c_clone = TreeBuildConfig {
+                property_mode: c.property_mode,
+                timeout_per_operation_ms: c.timeout_per_operation_ms,
+                yield_every_n_elements: c.yield_every_n_elements,
+                batch_size: c.batch_size,
+                max_depth: c.max_depth,
+                ui_settle_delay_ms: c.ui_settle_delay_ms,
+                format_output: c.format_output,
+                tree_output_format: c.tree_output_format,
+                tree_from_selector: c.tree_from_selector.clone(),
+                include_window_screenshot: c.include_window_screenshot,
+                include_monitor_screenshots: c.include_monitor_screenshots,
+                include_gemini_vision: None,
+                include_omniparser: None,
+                include_ocr: None,
+                include_browser_dom: None,
+            };
+            if matches!(output_format, TreeOutputFormat::VerboseJson) {
+                c_clone.format_output = Some(false);
+            } else if c_clone.format_output.is_none() {
+                c_clone.format_output = Some(true);
+            }
+            c_clone.into()
+        });
+
+        // Get UIA tree (always)
+        let result = self
+            .inner
+            .get_window_tree_result_async(pid, title.as_deref(), rust_config)
+            .await
+            .map_err(map_error)?;
+
+        let mut sdk_result = WindowTreeResult::from(result);
+
+        // If no vision options and not clustered format, return simple result
+        if !has_vision_options && !matches!(output_format, TreeOutputFormat::ClusteredYaml) {
+            if matches!(output_format, TreeOutputFormat::VerboseJson) {
+                sdk_result.formatted =
+                    Some(serde_json::to_string_pretty(&sdk_result.tree).unwrap_or_default());
+            }
+            return Ok(sdk_result);
+        }
+
+        // Build UIA bounds cache from formatted result
+        #[allow(clippy::type_complexity)]
+        let mut uia_bounds: HashMap<
+            u32,
+            (String, String, (f64, f64, f64, f64), Option<String>),
+        > = HashMap::new();
+        let uia_tree_result = self
+            .inner
+            .get_window_tree_result(pid, None, None)
+            .map_err(map_error)?;
+        let formatted_result = computeruse::format_ui_node_as_compact_yaml(&uia_tree_result.tree, 0);
+        for (idx, (role, name, bounds, selector)) in formatted_result.index_to_bounds {
+            uia_bounds.insert(idx, (role, name, bounds, selector));
+        }
+
+        // Build DOM bounds cache if requested
+        #[allow(clippy::type_complexity)]
+        let mut dom_bounds: HashMap<u32, (String, String, (f64, f64, f64, f64))> = HashMap::new();
+        if include_browser_dom && computeruse::is_browser_process(pid) {
+            if let Ok(dom_result) = self.capture_browser_dom(Some(100), Some(true)).await {
+                for (idx_str, entry) in dom_result.index_to_bounds {
+                    if let Ok(idx) = idx_str.parse::<u32>() {
+                        let bounds = (
+                            entry.bounds.x,
+                            entry.bounds.y,
+                            entry.bounds.width,
+                            entry.bounds.height,
+                        );
+                        dom_bounds.insert(idx, (entry.tag, entry.name, bounds));
+                    }
+                }
+            }
+        }
+
+        // Build Omniparser cache if requested
+        let mut omniparser_items: HashMap<u32, computeruse::OmniparserItem> = HashMap::new();
+        if include_omniparser {
+            if let Ok(omni_result) = self
+                .perform_omniparser_for_process(process.clone(), None, Some(true))
+                .await
+            {
+                for (idx_str, entry) in omni_result.index_to_bounds {
+                    if let Ok(idx) = idx_str.parse::<u32>() {
+                        omniparser_items.insert(
+                            idx,
+                            computeruse::OmniparserItem {
+                                label: entry.label.clone(),
+                                content: Some(entry.name.clone()),
+                                box_2d: Some([
+                                    entry.bounds.x,
+                                    entry.bounds.y,
+                                    entry.bounds.x + entry.bounds.width,
+                                    entry.bounds.y + entry.bounds.height,
+                                ]),
+                            },
+                        );
+                    }
+                }
+            }
+        }
+
+        // Build Gemini Vision cache if requested
+        let mut vision_items: HashMap<u32, computeruse::VisionElement> = HashMap::new();
+        if include_gemini_vision {
+            if let Ok(vision_result) = self
+                .perform_gemini_vision_for_process(process.clone(), Some(true))
+                .await
+            {
+                for (idx_str, entry) in vision_result.index_to_bounds {
+                    if let Ok(idx) = idx_str.parse::<u32>() {
+                        vision_items.insert(
+                            idx,
+                            computeruse::VisionElement {
+                                element_type: entry.element_type.clone(),
+                                content: Some(entry.name.clone()),
+                                description: None,
+                                box_2d: Some([
+                                    entry.bounds.x,
+                                    entry.bounds.y,
+                                    entry.bounds.x + entry.bounds.width,
+                                    entry.bounds.y + entry.bounds.height,
+                                ]),
+                                interactivity: None,
+                            },
+                        );
+                    }
+                }
+            }
+        }
+
+        // Build OCR cache if requested
+        #[allow(clippy::type_complexity)]
+        let mut ocr_bounds: HashMap<u32, (String, (f64, f64, f64, f64))> = HashMap::new();
+        if include_ocr {
+            if let Ok(ocr_result) = self
+                .perform_ocr_for_process(process.clone(), Some(true))
+                .await
+            {
+                for (idx_str, entry) in ocr_result.index_to_bounds {
+                    if let Ok(idx) = idx_str.parse::<u32>() {
+                        let bounds = (
+                            entry.bounds.x,
+                            entry.bounds.y,
+                            entry.bounds.width,
+                            entry.bounds.height,
+                        );
+                        ocr_bounds.insert(idx, (entry.text.clone(), bounds));
+                    }
+                }
+            }
+        }
+
+        // If ClusteredYaml format, use clustering
+        if matches!(output_format, TreeOutputFormat::ClusteredYaml) {
+            let clustered_result = computeruse::format_clustered_tree_from_caches(
+                &uia_bounds,
+                &dom_bounds,
+                &ocr_bounds,
+                &omniparser_items,
+                &vision_items,
+            );
+            sdk_result.formatted = Some(clustered_result.formatted);
+        } else {
+            // CompactYaml with vision - append vision trees to UIA tree
+            let mut combined = sdk_result.formatted.unwrap_or_default();
+            if !dom_bounds.is_empty() {
+                combined.push_str("\n\n# Browser DOM elements:\n");
+                for (idx, (tag, name, _)) in &dom_bounds {
+                    combined.push_str(&format!("#d{} [{}] {}\n", idx, tag, name));
+                }
+            }
+            if !omniparser_items.is_empty() {
+                combined.push_str("\n\n# Omniparser elements:\n");
+                for (idx, item) in &omniparser_items {
+                    combined.push_str(&format!(
+                        "#p{} [{}] {}\n",
+                        idx,
+                        item.label,
+                        item.content.as_deref().unwrap_or("")
+                    ));
+                }
+            }
+            if !vision_items.is_empty() {
+                combined.push_str("\n\n# Gemini Vision elements:\n");
+                for (idx, item) in &vision_items {
+                    combined.push_str(&format!(
+                        "#g{} [{}] {}\n",
+                        idx,
+                        item.element_type,
+                        item.content.as_deref().unwrap_or("")
+                    ));
+                }
+            }
+            if !ocr_bounds.is_empty() {
+                combined.push_str("\n\n# OCR elements:\n");
+                for (idx, (text, _)) in &ocr_bounds {
+                    combined.push_str(&format!("#o{} {}\n", idx, text));
+                }
+            }
+            sdk_result.formatted = Some(combined);
+        }
+
+        Ok(sdk_result)
+    }
+
+    // ============== NEW MONITOR METHODS ==============
+
+    /// (async) List all available monitors/displays.
+    ///
+    /// @returns {Promise<Array<Monitor>>} List of monitor information.
+    #[napi]
+    pub async fn list_monitors(&self) -> napi::Result<Vec<Monitor>> {
+        self.inner
+            .list_monitors()
+            .await
+            .map(|monitors| monitors.into_iter().map(Monitor::from).collect())
+            .map_err(map_error)
+    }
+
+    /// (async) Get the primary monitor.
+    ///
+    /// @returns {Promise<Monitor>} Primary monitor information.
+    #[napi]
+    pub async fn get_primary_monitor(&self) -> napi::Result<Monitor> {
+        self.inner
+            .get_primary_monitor()
+            .await
+            .map(Monitor::from)
+            .map_err(map_error)
+    }
+
+    /// (async) Get the monitor containing the currently focused window.
+    ///
+    /// @returns {Promise<Monitor>} Active monitor information.
+    #[napi]
+    pub async fn get_active_monitor(&self) -> napi::Result<Monitor> {
+        self.inner
+            .get_active_monitor()
+            .await
+            .map(Monitor::from)
+            .map_err(map_error)
+    }
+
+    /// (async) Get a monitor by its ID.
+    ///
+    /// @param {string} id - The monitor ID to find.
+    /// @returns {Promise<Monitor>} Monitor information.
+    #[napi]
+    pub async fn get_monitor_by_id(&self, id: String) -> napi::Result<Monitor> {
+        self.inner
+            .get_monitor_by_id(&id)
+            .await
+            .map(Monitor::from)
+            .map_err(map_error)
+    }
+
+    /// (async) Get a monitor by its name.
+    ///
+    /// @param {string} name - The monitor name to find.
+    /// @returns {Promise<Monitor>} Monitor information.
+    #[napi]
+    pub async fn get_monitor_by_name(&self, name: String) -> napi::Result<Monitor> {
+        self.inner
+            .get_monitor_by_name(&name)
+            .await
+            .map(Monitor::from)
+            .map_err(map_error)
+    }
+
+    /// (async) Capture a screenshot of a specific monitor.
+    ///
+    /// @param {Monitor} monitor - The monitor to capture.
+    /// @returns {Promise<ScreenshotResult>} The screenshot data.
+    #[napi]
+    pub async fn capture_monitor(&self, monitor: Monitor) -> napi::Result<ScreenshotResult> {
+        let rust_monitor = computeruse::Monitor {
+            id: monitor.id,
+            name: monitor.name,
+            is_primary: monitor.is_primary,
+            width: monitor.width,
+            height: monitor.height,
+            x: monitor.x,
+            y: monitor.y,
+            scale_factor: monitor.scale_factor,
+            work_area: None,
+        };
+        self.inner
+            .capture_monitor(&rust_monitor)
+            .await
+            .map(|r| ScreenshotResult {
+                width: r.width,
+                height: r.height,
+                image_data: r.image_data,
+                monitor: r.monitor.map(Monitor::from),
+            })
+            .map_err(map_error)
+    }
+
+    /// (async) Capture screenshots of all monitors.
+    ///
+    /// @returns {Promise<Array<{monitor: Monitor, screenshot: ScreenshotResult}>>} Array of monitor and screenshot pairs.
+    #[napi]
+    pub async fn capture_all_monitors(&self) -> napi::Result<Vec<MonitorScreenshotPair>> {
+        self.inner
+            .capture_all_monitors()
+            .await
+            .map(|results| {
+                results
+                    .into_iter()
+                    .map(|(monitor, screenshot)| MonitorScreenshotPair {
+                        monitor: Monitor::from(monitor),
+                        screenshot: ScreenshotResult {
+                            width: screenshot.width,
+                            height: screenshot.height,
+                            image_data: screenshot.image_data,
+                            monitor: screenshot.monitor.map(Monitor::from),
+                        },
+                    })
+                    .collect()
+            })
+            .map_err(map_error)
+    }
+
+    /// Capture a screenshot of a window by process name.
+    ///
+    /// Finds the first window matching the given process name and captures its screenshot.
+    /// Process name matching is case-insensitive and uses substring matching.
+    ///
+    /// @param {string} process - Process name to match (e.g., "chrome", "notepad", "code")
+    /// @returns {ScreenshotResult} The screenshot data.
+    #[napi(js_name = "captureWindowByProcess")]
+    pub fn capture_window_by_process(&self, process: String) -> napi::Result<ScreenshotResult> {
+        self.inner
+            .capture_window_by_process(&process)
+            .map(|r| ScreenshotResult {
+                width: r.width,
+                height: r.height,
+                image_data: r.image_data,
+                monitor: r.monitor.map(Monitor::from),
+            })
+            .map_err(map_error)
+    }
+
+    /// (async) Captures a screenshot. Three modes:
+    /// 1. Element mode: provide process + selector to capture specific element
+    /// 2. Window mode: provide process only to capture entire window
+    /// 3. Monitor mode: provide process + entireMonitor=true to capture the monitor where the window is located
+    ///
+    /// @param {string} process - Process name to match (e.g., "chrome", "notepad", "code")
+    /// @param {string} [selector] - Optional selector to capture a specific element within the process
+    /// @param {boolean} [entireMonitor=false] - If true, captures the entire monitor containing the window
+    /// @param {number} [timeoutMs=10000] - Timeout in milliseconds for finding the element
+    /// @returns {Promise<ScreenshotResult>} The screenshot data.
+    #[napi(js_name = "captureScreenshot")]
+    pub async fn capture_screenshot(
+        &self,
+        process: String,
+        selector: Option<String>,
+        entire_monitor: Option<bool>,
+        timeout_ms: Option<f64>,
+    ) -> napi::Result<ScreenshotResult> {
+        use std::time::Duration;
+
+        let entire_monitor = entire_monitor.unwrap_or(false);
+        let timeout = Duration::from_millis(timeout_ms.unwrap_or(10000.0) as u64);
+
+        // Build the full selector string like MCP does
+        let full_selector = if let Some(sel) = &selector {
+            if sel.is_empty() {
+                format!("process:{}", process)
+            } else {
+                format!("process:{} >> {}", process, sel)
+            }
+        } else {
+            format!("process:{}", process)
+        };
+
+        // Create locator and find element
+        let sel_rust: computeruse::selector::Selector = full_selector.as_str().into();
+        let locator = self.inner.locator(sel_rust);
+        let element = locator.first(Some(timeout)).await.map_err(map_error)?;
+
+        if entire_monitor {
+            // Monitor mode: get element's monitor and capture it
+            let monitor = element.monitor().map_err(map_error)?;
+            let screenshot = monitor.capture(&self.inner).await.map_err(map_error)?;
+            Ok(ScreenshotResult {
+                width: screenshot.width,
+                height: screenshot.height,
+                image_data: screenshot.image_data,
+                monitor: Some(Monitor::from(monitor)),
+            })
+        } else {
+            // Element/Window mode: capture the element directly
+            let screenshot = element.capture().map_err(map_error)?;
+            Ok(ScreenshotResult {
+                width: screenshot.width,
+                height: screenshot.height,
+                image_data: screenshot.image_data,
+                monitor: screenshot.monitor.map(Monitor::from),
+            })
+        }
+    }
+
+    // ============== SCREENSHOT UTILITIES ==============
+
+    /// Convert a screenshot to PNG bytes.
+    /// Converts BGRA to RGBA and encodes as PNG format.
+    ///
+    /// @param {ScreenshotResult} screenshot - The screenshot to convert.
+    /// @returns {Buffer} PNG-encoded bytes.
+    #[napi(js_name = "screenshotToPng")]
+    pub fn screenshot_to_png(&self, screenshot: ScreenshotResult) -> napi::Result<Vec<u8>> {
+        screenshot
+            .to_inner()
+            .to_png()
+            .map_err(|e| napi::Error::from_reason(e.to_string()))
+    }
+
+    /// Convert a screenshot to PNG bytes with resizing.
+    /// If the image exceeds maxDimension in either width or height,
+    /// it will be resized while maintaining aspect ratio.
+    ///
+    /// @param {ScreenshotResult} screenshot - The screenshot to convert.
+    /// @param {number} [maxDimension] - Maximum width or height. Defaults to 1920.
+    /// @returns {Buffer} PNG-encoded bytes (potentially resized).
+    #[napi(js_name = "screenshotToPngResized")]
+    pub fn screenshot_to_png_resized(
+        &self,
+        screenshot: ScreenshotResult,
+        max_dimension: Option<u32>,
+    ) -> napi::Result<Vec<u8>> {
+        screenshot
+            .to_inner()
+            .to_png_resized(max_dimension)
+            .map_err(|e| napi::Error::from_reason(e.to_string()))
+    }
+
+    /// Convert a screenshot to base64-encoded PNG string.
+    /// Useful for embedding in JSON responses or passing to LLMs.
+    ///
+    /// @param {ScreenshotResult} screenshot - The screenshot to convert.
+    /// @returns {string} Base64-encoded PNG string.
+    #[napi(js_name = "screenshotToBase64Png")]
+    pub fn screenshot_to_base64_png(&self, screenshot: ScreenshotResult) -> napi::Result<String> {
+        screenshot
+            .to_inner()
+            .to_base64_png()
+            .map_err(|e| napi::Error::from_reason(e.to_string()))
+    }
+
+    /// Convert a screenshot to base64-encoded PNG string with resizing.
+    /// If the image exceeds maxDimension in either width or height,
+    /// it will be resized while maintaining aspect ratio.
+    ///
+    /// @param {ScreenshotResult} screenshot - The screenshot to convert.
+    /// @param {number} [maxDimension] - Maximum width or height. Defaults to 1920.
+    /// @returns {string} Base64-encoded PNG string (potentially resized).
+    #[napi(js_name = "screenshotToBase64PngResized")]
+    pub fn screenshot_to_base64_png_resized(
+        &self,
+        screenshot: ScreenshotResult,
+        max_dimension: Option<u32>,
+    ) -> napi::Result<String> {
+        screenshot
+            .to_inner()
+            .to_base64_png_resized(max_dimension)
+            .map_err(|e| napi::Error::from_reason(e.to_string()))
+    }
+
+    /// Get the dimensions a screenshot would have after resizing.
+    ///
+    /// @param {ScreenshotResult} screenshot - The screenshot to check.
+    /// @param {number} maxDimension - Maximum width or height.
+    /// @returns {ResizedDimensions} Object with width and height after resize.
+    #[napi(js_name = "screenshotResizedDimensions")]
+    pub fn screenshot_resized_dimensions(
+        &self,
+        screenshot: ScreenshotResult,
+        max_dimension: u32,
+    ) -> ResizedDimensions {
+        let (width, height) = screenshot.to_inner().resized_dimensions(max_dimension);
+        ResizedDimensions { width, height }
+    }
+
+    /// (async) Get all window elements for a given application name.
+    ///
+    /// @param {string} name - The name of the application whose windows will be retrieved.
+    /// @returns {Promise<Array<Element>>} A list of window elements belonging to the application.
+    #[napi]
+    pub async fn windows_for_application(&self, name: String) -> napi::Result<Vec<Element>> {
+        self.inner
+            .windows_for_application(&name)
+            .await
+            .map(|windows| windows.into_iter().map(Element::from).collect())
+            .map_err(map_error)
+    }
+
+    // ============== ADDITIONAL MISSING METHODS ==============
+
+    /// (async) Get the UI tree for all open applications in parallel.
+    ///
+    /// @returns {Promise<Array<UINode>>} List of UI trees for all applications.
+    #[napi]
+    pub async fn get_all_applications_tree(&self) -> napi::Result<Vec<UINode>> {
+        self.inner
+            .get_all_applications_tree()
+            .await
+            .map(|trees| trees.into_iter().map(UINode::from).collect())
+            .map_err(map_error)
+    }
+
+    /// (async) Press a key globally.
+    ///
+    /// @param {string} key - The key to press (e.g., "Enter", "Ctrl+C", "F1").
+    /// @param {string} [process] - Process name for window screenshot capture. If provided, enables window screenshots.
+    /// @param {boolean} [includeWindowScreenshot=true] - Whether to capture window screenshot (requires process).
+    /// @param {boolean} [includeMonitorScreenshots=false] - Whether to capture monitor screenshots after key press.
+    #[napi]
+    pub async fn press_key(
+        &self,
+        key: String,
+        process: Option<String>,
+        include_window_screenshot: Option<bool>,
+        include_monitor_screenshots: Option<bool>,
+    ) -> napi::Result<()> {
+        // Normalize key to ensure curly brace format (e.g., "Enter" -> "{Enter}")
+        let normalized_key = normalize_key(&key);
+        tracing::debug!(
+            "[TS SDK] desktop.press_key: normalized key: {} -> {}",
+            key,
+            normalized_key
+        );
+        let result = self
+            .inner
+            .press_key(&normalized_key)
+            .await
+            .map_err(map_error);
+
+        // Get PID from process name if provided
+        let pid = process
+            .as_ref()
+            .and_then(|p| find_pid_for_process(&self.inner, p).ok());
+
+        // Capture screenshots if requested
+        let _screenshots = capture_screenshots(
+            &self.inner,
+            pid,
+            include_window_screenshot.unwrap_or(true) && pid.is_some(),
+            include_monitor_screenshots.unwrap_or(false),
+            "pressKey",
+        );
+
+        result
+    }
+
+    /// (async) Execute JavaScript in a browser tab.
+    /// Finds the browser window by process name and executes the script.
+    ///
+    /// @param {string} script - The JavaScript code to execute in browser context.
+    /// @param {string} process - Process name to scope the browser window (e.g., 'chrome', 'msedge'). Required.
+    /// @param {number} [timeoutMs=10000] - Timeout in milliseconds for finding the browser window.
+    /// @returns {Promise<string>} The result of script execution.
+    #[napi]
+    pub async fn execute_browser_script(
+        &self,
+        script: String,
+        process: String,
+        timeout_ms: Option<f64>,
+    ) -> napi::Result<String> {
+        use std::time::Duration;
+
+        let timeout = Duration::from_millis(timeout_ms.unwrap_or(10000.0) as u64);
+        let selector_str = format!("process:{}", process);
+        let sel: computeruse::selector::Selector = selector_str.as_str().into();
+        let locator = self.inner.locator(sel);
+        let element = locator.first(Some(timeout)).await.map_err(map_error)?;
+        element
+            .execute_browser_script(&script)
+            .await
+            .map_err(map_error)
+    }
+
+    /// (async) Close a browser tab safely.
+    ///
+    /// This method can identify the tab to close by:
+    /// - tabId: Close a specific tab by its Chrome tab ID
+    /// - url: Find and close a tab matching this URL (partial match supported)
+    /// - title: Find and close a tab matching this title (case-insensitive partial match)
+    /// - If none provided, closes the currently active tab
+    ///
+    /// Returns information about the closed tab for verification.
+    /// Returns null if no browser extension is connected or tab couldn't be found.
+    ///
+    /// Safety:
+    /// - Will NOT close protected browser pages (chrome://, edge://, about:, etc.)
+    /// - Returns the closed tab's URL/title so you can verify the correct tab was closed
+    ///
+    /// @param {number} [tabId] - Specific Chrome tab ID to close.
+    /// @param {string} [url] - URL to match (partial match supported).
+    /// @param {string} [title] - Title to match (case-insensitive partial match).
+    /// @returns {Promise<CloseTabResult | null>} Info about closed tab, or null if no extension/tab found.
+    ///
+    /// @example
+    /// // Close by URL
+    /// const result = await desktop.closeTab({ url: "example.com" });
+    ///
+    /// @example
+    /// // Close by title
+    /// const result = await desktop.closeTab({ title: "My Page" });
+    ///
+    /// @example
+    /// // Close active tab
+    /// const result = await desktop.closeTab();
+    #[napi]
+    pub async fn close_tab(
+        &self,
+        options: Option<crate::types::CloseTabOptions>,
+    ) -> napi::Result<Option<crate::types::CloseTabResult>> {
+        let opts = options.unwrap_or_default();
+        self.inner
+            .close_tab(opts.tab_id, opts.url.as_deref(), opts.title.as_deref())
+            .await
+            .map(|opt| opt.map(crate::types::CloseTabResult::from))
+            .map_err(map_error)
+    }
+    /// (async) Delay execution for a specified number of milliseconds.
+    /// Useful for waiting between actions to ensure UI stability.
+    ///
+    /// @param {number} delayMs - Delay in milliseconds.
+    /// @returns {Promise<void>}
+    #[napi]
+    pub async fn delay(&self, delay_ms: u32) -> napi::Result<()> {
+        self.inner.delay(delay_ms as u64).await.map_err(map_error)
+    }
+
+    /// Navigate to a URL in a browser.
+    /// This is the recommended method for browser navigation - more reliable than
+    /// manually manipulating the address bar with keyboard/mouse actions.
+    ///
+    /// @param {string} url - URL to navigate to
+    /// @param {string | null} browser - Optional browser name ('Chrome', 'Firefox', 'Edge', 'Brave', 'Opera', 'Vivaldi', or 'Default')
+    /// @param {boolean} [includeWindowScreenshot=true] - Whether to capture window screenshot after navigation
+    /// @param {boolean} [includeMonitorScreenshots=false] - Whether to capture monitor screenshots after navigation
+    /// @returns {Promise<Element>} The browser window element
+    #[napi]
+    pub fn navigate_browser(
+        &self,
+        url: String,
+        browser: Option<String>,
+        include_window_screenshot: Option<bool>,
+        include_monitor_screenshots: Option<bool>,
+    ) -> napi::Result<Element> {
+        let browser_enum = browser.map(|b| match b.as_str() {
+            "Chrome" => computeruse::Browser::Chrome,
+            "Firefox" => computeruse::Browser::Firefox,
+            "Edge" => computeruse::Browser::Edge,
+            "Brave" => computeruse::Browser::Brave,
+            "Opera" => computeruse::Browser::Opera,
+            "Vivaldi" => computeruse::Browser::Vivaldi,
+            "Default" => computeruse::Browser::Default,
+            custom => computeruse::Browser::Custom(custom.to_string()),
+        });
+
+        let element = self.inner.open_url(&url, browser_enum).map_err(map_error)?;
+
+        // Capture screenshots if enabled (window default: true, monitor default: false)
+        let _screenshots = capture_screenshots(
+            &self.inner,
+            element.process_id().ok(),
+            include_window_screenshot.unwrap_or(true),
+            include_monitor_screenshots.unwrap_or(false),
+            "navigateBrowser",
+        );
+
+        Ok(Element { inner: element })
+    }
+
+    /// (async) Set the zoom level to a specific percentage.
+    ///
+    /// @param {number} percentage - The zoom percentage (e.g., 100 for 100%, 150 for 150%, 50 for 50%).
+    #[napi]
+    pub async fn set_zoom(&self, percentage: u32) -> napi::Result<()> {
+        self.inner.set_zoom(percentage).await.map_err(map_error)
+    }
+
+    /// (async) Run Gemini Computer Use agentic loop.
+    ///
+    /// Provide a goal and target process, and this will autonomously take actions
+    /// (click, type, scroll, etc.) until the goal is achieved or max_steps is reached.
+    /// Uses Gemini's vision model to analyze screenshots and decide actions.
+    ///
+    /// @param {string} process - Process name of the target application (e.g., "chrome", "notepad")
+    /// @param {string} goal - What to achieve (e.g., "Open Notepad and type Hello World")
+    /// @param {number} [maxSteps=20] - Maximum number of steps before stopping
+    /// @param {function} [onStep] - Optional callback invoked after each step with step details
+    /// @returns {Promise<ComputerUseResult>} Result with status, steps executed, and history
+    #[napi]
+    pub async fn gemini_computer_use(
+        &self,
+        process: String,
+        goal: String,
+        max_steps: Option<u32>,
+        #[napi(ts_arg_type = "((err: null | Error, step: ComputerUseStep) => void) | undefined")]
+        on_step: Option<ThreadsafeFunction<ComputerUseStep>>,
+    ) -> napi::Result<ComputerUseResult> {
+        // Create progress callback if onStep is provided
+        #[allow(clippy::type_complexity)]
+        let progress_callback: Option<
+            Box<dyn Fn(&computeruse::ComputerUseStep) + Send + Sync>,
+        > = on_step.map(|tsfn| {
+            let tsfn = Arc::new(tsfn);
+            Box::new(move |step: &computeruse::ComputerUseStep| {
+                let js_step = ComputerUseStep::from(step.clone());
+                tsfn.call(Ok(js_step), ThreadsafeFunctionCallMode::NonBlocking);
+            }) as Box<dyn Fn(&computeruse::ComputerUseStep) + Send + Sync>
+        });
+
+        self.inner
+            .gemini_computer_use(&process, &goal, max_steps, progress_callback)
+            .await
+            .map(ComputerUseResult::from)
+            .map_err(|e| napi::Error::from_reason(e.to_string()))
+    }
+
+    /// Stop all currently executing operations.
+    ///
+    /// This cancels the internal cancellation token, which will cause any
+    /// operations that check `isCancelled()` to abort. After calling this,
+    /// you should create a new Desktop instance to start fresh.
+    #[napi]
+    pub fn stop_execution(&self) {
+        self.inner.stop_execution();
+    }
+
+    /// Check if execution has been cancelled.
+    ///
+    /// Returns `true` if `stopExecution()` has been called.
+    /// Long-running operations should periodically check this and abort if true.
+    #[napi]
+    pub fn is_cancelled(&self) -> bool {
+        self.inner.is_cancelled()
+    }
+
+    /// Stop all active highlight overlays globally.
+    ///
+    /// This finds and destroys all highlight overlay windows that were created
+    /// by `element.highlight()`. Useful for cleaning up highlights without
+    /// needing to track individual HighlightHandle objects.
+    ///
+    /// @returns {number} The number of highlights that were stopped.
+    #[napi]
+    pub fn stop_highlighting(&self) -> u32 {
+        #[cfg(target_os = "windows")]
+        {
+            computeruse::stop_all_highlights() as u32
+        }
+        #[cfg(not(target_os = "windows"))]
+        {
+            // Not implemented for other platforms yet
+            0
+        }
+    }
+
+    /// Show inspect overlay with indexed elements for visual debugging.
+    ///
+    /// Displays a transparent overlay window with colored rectangles around UI elements,
+    /// showing their index numbers for click targeting. Use `hideInspectOverlay()` to remove.
+    ///
+    /// @param {InspectElement[]} elements - Array of elements to highlight with their bounds.
+    /// @param {object} windowBounds - The window bounds {x, y, width, height} to constrain the overlay.
+    /// @param {OverlayDisplayMode} [displayMode='Index'] - What to show in labels: 'Index', 'Role', 'Name', etc.
+    #[napi]
+    #[cfg(target_os = "windows")]
+    pub fn show_inspect_overlay(
+        &self,
+        elements: Vec<crate::types::InspectElement>,
+        window_bounds: crate::types::Bounds,
+        display_mode: Option<crate::types::OverlayDisplayMode>,
+    ) -> napi::Result<()> {
+        let core_elements: Vec<computeruse::InspectElement> =
+            elements.into_iter().map(|e| e.into()).collect();
+        let core_bounds = (
+            window_bounds.x as i32,
+            window_bounds.y as i32,
+            window_bounds.width as i32,
+            window_bounds.height as i32,
+        );
+        let core_mode = display_mode
+            .map(|m| m.into())
+            .unwrap_or(computeruse::OverlayDisplayMode::Index);
+
+        computeruse::show_inspect_overlay(core_elements, core_bounds, core_mode)
+            .map(|_handle| ()) // Discard handle - use hideInspectOverlay to close
+            .map_err(|e| napi::Error::from_reason(e.to_string()))
+    }
+
+    /// Show inspect overlay (non-Windows stub).
+    #[napi]
+    #[cfg(not(target_os = "windows"))]
+    pub fn show_inspect_overlay(
+        &self,
+        _elements: Vec<crate::types::InspectElement>,
+        _window_bounds: crate::types::Bounds,
+        _display_mode: Option<crate::types::OverlayDisplayMode>,
+    ) -> napi::Result<()> {
+        // Not implemented for other platforms yet
+        Ok(())
+    }
+
+    /// Hide any active inspect overlay.
+    ///
+    /// This hides the visual overlay that was shown via `showInspectOverlay()`.
+    /// Can be called from any thread.
+    #[napi]
+    pub fn hide_inspect_overlay(&self) {
+        #[cfg(target_os = "windows")]
+        {
+            computeruse::hide_inspect_overlay();
+        }
+        #[cfg(not(target_os = "windows"))]
+        {
+            // Not implemented for other platforms yet
+        }
+    }
+
+    // ============== ELEMENT VERIFICATION ==============
+
+    /// Verify that an element matching the selector exists within the same application as the scope element.
+    ///
+    /// This is used for post-action verification - checking that an expected element appeared after
+    /// performing an action (e.g., a success dialog after clicking submit).
+    ///
+    /// @param {Element} scopeElement - The element to get the application scope from (typically the element the action was performed on)
+    /// @param {string} selector - The selector string to search for
+    /// @param {number} [timeoutMs=2000] - How long to wait for the element to appear in milliseconds
+    /// @returns {Element} The found element if verification passes
+    /// @throws Error if the element is not found within the timeout
+    #[napi]
+    pub async fn verify_element_exists(
+        &self,
+        scope_element: &crate::Element,
+        selector: String,
+        timeout_ms: Option<u32>,
+    ) -> napi::Result<crate::Element> {
+        let timeout = timeout_ms.unwrap_or(2000) as u64;
+        let found = self
+            .inner
+            .verify_element_exists(&scope_element.inner, &selector, timeout)
+            .await
+            .map_err(map_error)?;
+        Ok(crate::Element { inner: found })
+    }
+
+    /// Verify that an element matching the selector does NOT exist within the same application as the scope element.
+    ///
+    /// This is used for post-action verification - checking that an element disappeared after
+    /// performing an action (e.g., a modal dialog closed after clicking OK).
+    ///
+    /// @param {Element} scopeElement - The element to get the application scope from (typically the element the action was performed on)
+    /// @param {string} selector - The selector string that should NOT be found
+    /// @param {number} [timeoutMs=2000] - How long to wait/check that the element doesn't appear in milliseconds
+    /// @returns {void}
+    /// @throws Error if the element IS found (meaning verification failed)
+    #[napi]
+    pub async fn verify_element_not_exists(
+        &self,
+        scope_element: &crate::Element,
+        selector: String,
+        timeout_ms: Option<u32>,
+    ) -> napi::Result<()> {
+        let timeout = timeout_ms.unwrap_or(2000) as u64;
+        self.inner
+            .verify_element_not_exists(&scope_element.inner, &selector, timeout)
+            .await
+            .map_err(map_error)
+    }
+}