windows-chat-verification

name	windows-chat-verification
description	Verify real Windows chat-app workflows through DesktopManager MCP. Use when an MCP-capable assistant should open a desktop chat app, inspect its window and controls, type into the prompt area, send a message, and verify the visible result with screenshots and structural or app-owned evidence.

Windows Chat Verification

Use this skill when the task is to verify a real Windows chat application through DesktopManager, especially IntelligenceX Chat and similar WebView2 or Chromium-style shells.

Golden Path

1. Make sure the DesktopManager MCP server is available through this plugin.
   • The plugin starts DesktopManager through ./scripts/start-desktopmanager-mcp.ps1.
   • Build DesktopManager.Cli first if the repo has not produced a CLI host yet.
   • Keep the default server read-only unless the verification needs clicks, typing, or layout changes.
2. Enable only the minimum mutation scope you need.
   • Set DESKTOPMANAGER_MCP_ALLOW_MUTATIONS=1 before starting the MCP client session when the workflow needs live interaction.
   • Use DESKTOPMANAGER_MCP_ALLOW_PROCESS=IntelligenceX.Chat.App;DesktopManager.TestApp or another narrow process list when you can.
   • Set DESKTOPMANAGER_MCP_ALLOW_FOREGROUND_INPUT=1 only when a zero-handle UIA editor really needs focused fallback text or key input.
3. Start with inspection, not clicks.
   • Read desktop://windows/visible, desktop://windows/active, and desktop://monitors.
   • Use list_windows, get_active_window, and window_exists before assuming the target app is ready.
   • Capture screenshot_window or screenshot_desktop when visual proof matters.
4. Launch or attach to the target chat app.
   • For IntelligenceX Chat in a sibling repo, use ./scripts/start-intelligencex-chat.ps1 if available.
   • Otherwise use launch_process or launch_and_wait_for_window.
   • Prefer wait_for_window over blind retries.
5. Inspect controls before typing.
   • Use list_window_controls first.
   • For WebView2 or Chromium-style surfaces, use diagnose_window_controls when expected edit controls are missing or unstable.
   • Prefer saved control targets when the same prompt area or send button will be reused.
6. Interact with the smallest safe target.
   • Prefer click_control, set_control_text, and send_control_keys when the app exposes usable controls.
   • If the chat editor is a zero-handle UIA surface, use explicit foreground fallback only when the server was started with --allow-foreground-input.
   • If structure still fails, switch to screenshot-assisted targeting with get_window_geometry, save_window_target, and click_window_point.
   • For IntelligenceX.Chat, prefer the app-owned automation contract through ./scripts/invoke-intelligencex-chat-automation.ps1 before spending time on coordinate fallbacks.
7. Verify the result from more than one angle.
   • Capture an after screenshot.
   • Re-read the relevant controls or window text when possible.
   • Confirm the visible transcript or status change instead of trusting the input acknowledgement alone.
   • When IntelligenceX.Chat runs in automation mode, confirm wait_for_idle, get_transcript_tail, or list_conversations from the app-owned pipe.

IntelligenceX Notes

• IntelligenceX Chat currently uses a WebView2 shell with a composer textarea and send button in the web UI.
• In the sibling repo, the current shell template exposes textarea#prompt and button#btnSend.
• Treat that as helpful context, not a guarantee that UI Automation will expose the same shape on every run.
• IntelligenceX.Chat now also exposes an opt-in app-owned automation pipe in automation mode. That is the preferred verification path when the visible WebView surface is opaque to UIA.
• For IntelligenceX verification, prefer this flow:
  1. launch the app in automation mode or attach to an existing automation-mode instance
  2. capture the real window with DesktopManager so you still have visual evidence
  3. call ./scripts/invoke-intelligencex-chat-automation.ps1 -Command new_conversation -LaunchIfNeeded
  4. call ./scripts/invoke-intelligencex-chat-automation.ps1 -Command send_prompt -Text "Reply with only the single word PONG."
  5. call ./scripts/invoke-intelligencex-chat-automation.ps1 -Command wait_for_idle
  6. call ./scripts/invoke-intelligencex-chat-automation.ps1 -Command get_transcript_tail -Count 2
  7. confirm the latest assistant item matches the expected reply
  8. keep DesktopManager screenshots and window/control diagnostics alongside the app-owned verification

Decision Rules

• Prefer inspection-first MCP workflows over ad-hoc shell automation.
• Prefer launch_and_wait_for_window when the very next step depends on the real app window.
• Prefer UIA or control-aware actions before coordinate clicks.
• Prefer diagnose_window_controls before escalating to foreground fallback on modern apps.
• Prefer screenshot-assisted window targets over raw pixel guesses when structure is weak.
• Prefer captureAfter for mutating actions that matter.
• Prefer app-owned contracts over brittle coordinate replay when the target app explicitly provides one.
• If the server blocks mutating tools, explain that the MCP session is still in read-only mode instead of pretending the selector failed.

Useful Environment Variables

• DESKTOPMANAGER_REPO_ROOT
• DESKTOPMANAGER_MCP_EXE
• DESKTOPMANAGER_MCP_DLL
• DESKTOPMANAGER_MCP_ALLOW_MUTATIONS
• DESKTOPMANAGER_MCP_ALLOW_PROCESS
• DESKTOPMANAGER_MCP_DENY_PROCESS
• DESKTOPMANAGER_MCP_ALLOW_FOREGROUND_INPUT
• DESKTOPMANAGER_MCP_DRY_RUN
• INTELLIGENCEX_REPO_ROOT
• IXCHAT_AUTOMATION_MODE

Reference Files

• Docs/DesktopManager.Mcp.md
• .agents/skills/desktopmanager-operator/SKILL.md
• plugins/desktopmanager-windows-verification/scripts/start-desktopmanager-mcp.ps1
• plugins/desktopmanager-windows-verification/scripts/start-intelligencex-chat.ps1
• plugins/desktopmanager-windows-verification/scripts/invoke-intelligencex-chat-automation.ps1