feat: add UI Automation 'find' command for true textual vision

This commit is contained in:
Gemini AI
2025-12-14 02:31:53 +04:00
Unverified
parent ce4f0036c2
commit fb2991ba1a
2 changed files with 65 additions and 6 deletions

View File

@@ -9,6 +9,39 @@ param(
# Load required assemblies
Add-Type -AssemblyName System.Windows.Forms
Add-Type -AssemblyName System.Drawing
Add-Type -AssemblyName UIAutomationClient
Add-Type -AssemblyName UIAutomationTypes
# ... (Previous code remains) ...
switch ($Command.ToLower()) {
# ... (Previous cases remain) ...
"find" {
if ($Params.Count -lt 1) { Write-Error "Usage: find 'Name'"; exit 1 }
$targetName = $Params -join " "
Write-Host "Searching for UI Element: '$targetName'..."
$root = [System.Windows.Automation.AutomationElement]::RootElement
$cond = New-Object System.Windows.Automation.PropertyCondition([System.Windows.Automation.AutomationElement]::NameProperty, $targetName)
# Try finding directly (fast)
$element = $root.FindFirst([System.Windows.Automation.TreeScope]::Descendants, $cond)
if ($element) {
$rect = $element.Current.BoundingRectangle
$centerX = [int]($rect.X + ($rect.Width / 2))
$centerY = [int]($rect.Y + ($rect.Height / 2))
Write-Host "Found '$targetName' at ($centerX, $centerY)"
Write-Host "Action: mouse $centerX $centerY"
} else {
Write-Host "Element '$targetName' not found visible on desktop."
}
}
"apps" {
# ... (Rest remains) ...
# C# P/Invoke for advanced Input (SendInput is more reliable than SendKeys)
$code = @"
@@ -97,6 +130,31 @@ switch ($Command.ToLower()) {
Write-Host "Screenshot saved to $fullPath"
}
"find" {
if ($Params.Count -lt 1) { Write-Error "Usage: find 'Name'"; exit 1 }
$targetName = $Params -join " "
Write-Host "Searching for UI Element: '$targetName'..."
$root = [System.Windows.Automation.AutomationElement]::RootElement
$cond = New-Object System.Windows.Automation.PropertyCondition([System.Windows.Automation.AutomationElement]::NameProperty, $targetName)
# Try finding directly (fast)
$element = $root.FindFirst([System.Windows.Automation.TreeScope]::Descendants, $cond)
if ($element) {
$rect = $element.Current.BoundingRectangle
$centerX = [int]($rect.X + ($rect.Width / 2))
$centerY = [int]($rect.Y + ($rect.Height / 2))
Write-Host "Found '$targetName' at ($centerX, $centerY)"
# Auto-selection support return format
Write-Host "COORD:$centerX,$centerY"
} else {
Write-Host "Element '$targetName' not found visible on desktop."
}
}
"apps" {
# List all processes with a window title (visible apps)
$apps = Get-Process | Where-Object { $_.MainWindowTitle -ne "" } | Select-Object Id, MainWindowTitle

View File

@@ -491,14 +491,15 @@ Use it to control the mouse, keyboard, and "see" the system.
## 👁️ VISION & BLINDNESS PROTOCOL:
You are a TEXT-BASED intelligence. You CANNOT see images/screenshots you take.
- **\`input.ps1 screenshot\`**: Creates an image for the **USER** to see. You learn NOTHING from this.
- **\`input.ps1 find "Name"\`**: **TRUE VISION**. Finds a UI element (button/window) by text and tells you where it is.
- **\`input.ps1 apps\`**: Your "Eyes" for windows. Returns TEXT list of open apps.
- **\`input.ps1 screen\`**: Your "Eyes" for geometry. Returns TEXT resolution (e.g. 1920x1080).
- **\`input.ps1 screen\`**: Your "Eyes" for geometry. Returns TEXT resolution.
### 📐 THE LAW OF COORDINATES:
Since you cannot see buttons, you MUST calculate them using \`screen\` dimensions.
1. Run \`powershell bin/input.ps1 screen\`.
2. Get Output: \`Width x Height\` (e.g. 1920 x 1200).
### 📐 THE LAW OF ACCURACY:
1. **FIND FIRST**: If you need to click a button, SEARCH FOR IT.
- \`powershell bin/input.ps1 find "Start"\` -> Returns "Found at (30, 1190)".
- **THEN** use those coordinates to click.
2. **FALLBACK**: Only calculate coordinates manually if \`find\` fails.
3. **Start Menu Logic:** Bottom-Left corner.
- X = 0 to 50
- Y = Height - 10 (e.g. 1190).