letta-ai/remote-desktop-testing-windows

Remote Desktop Testing on Windows

Overview

Generic Daytona Windows sandbox driver. It creates (or reuses) a Windows sandbox whose desktop is streamed over noVNC, and provides the primitives for visual testing of any GUI app: sync code, run PowerShell, launch GUI programs in the interactive desktop session, screenshot, record video, and mint a clickable desktop link a teammate can open to drive the app.

When to Use

• Demo or visually verify a GUI app (Electron, browser, native Win32) on Windows without a local Windows machine
• Produce a screenshot + video + clickable live-desktop link for a teammate
• Debug Windows-specific desktop behavior of a cross-platform app
• Run throwaway desktop sessions that are fully disposable (stop with cleanup)

Usage

Run the bundled script with Node from the skill directory:

node scripts/remote-desktop.mjs <command> [options]

Credentials: DAYTONA_API_KEY (and optionally DAYTONA_API_URL/DAYTONA_TARGET) from the environment or a dotenv file — ./.env by default, override with --env-path <file>. The script auto-installs @daytona/sdk under ~/.letta/skill-state/remote-desktop-testing-windows/ on first use and remembers the active sandbox id there, so later commands don't need --sandbox.

Typical flow

# 1. Start (or reuse) a Windows sandbox (default snapshot: "windows", the stock Daytona one)
node scripts/remote-desktop.mjs start

# 2. (Optional) push a local project to C:\Users\Public\remote-desktop\workspace
node scripts/remote-desktop.mjs sync --project-path ~/repos/my-app --sync-mode working_tree

# 3. Set up / build with PowerShell commands
node scripts/remote-desktop.mjs shell \
  --command "Set-Location C:\Users\Public\remote-desktop\workspace; npm install"

# 4. Launch the GUI app in the interactive desktop session and verify its window appeared
node scripts/remote-desktop.mjs launch \
  --command "my-app.exe --some-flag" --label my-app --wait-window "My App"

# 5. Deliverables: screenshot, video, and a clickable live-desktop link
node scripts/remote-desktop.mjs screenshot --output /tmp/demo.png
node scripts/remote-desktop.mjs record --output /tmp/demo.mp4 --mp4 --duration 12
node scripts/remote-desktop.mjs preview

Always confirm the result visually (open the screenshot, or check windows / --wait-window) before reporting success — never report success from process existence alone. When sharing with a teammate, send the desktopUrl from preview (signed noVNC link, expires ~6h; re-run preview to refresh). Pass --port <n> to preview to also mint a signed URL for an app's HTTP port.

Commands

| Command | What it does | | --- | --- | | start | Create/reuse a sandbox and start the desktop stream. --fresh forces a new one. | | sync | Sync --project-path to C:\Users\Public\remote-desktop\workspace (git_archive = committed HEAD, working_tree = uncommitted too; robocopy /MIR preserving remote node_modules). | | shell --command PS | Run PowerShell. --timeout <s> for long ones. | | launch --command CMD | Launch a GUI command in the interactive desktop session via schtasks /IT (logged to C:\Users\Public\remote-desktop\launch-<label>.log). --wait-window <regex> verifies a matching window title appears. | | windows | List visible desktop windows. | | screenshot | Save a PNG of the full desktop locally. | | preview | Print signed desktopUrl (noVNC), plus appUrl if --port given. | | record | Record the live desktop via Playwright + local Chrome (--duration <s>, --mp4). | | recording-start/stop/download | Native Daytona recording — unreliable on Windows (recordings can silently vanish); prefer record. | | snapshot --snapshot <name> | Bake the sandbox into a snapshot (Daytona snapshots Windows from STOPPED; stops, snapshots, restarts). | | cleanup | Stop the active sandbox. |

Snapshots

Cold Windows boxes are slow to set up. After installing your toolchain/deps once, bake a prepared snapshot so future start --snapshot <name> runs skip setup:

node scripts/remote-desktop.mjs snapshot --snapshot my-app-windows-dev-v1

Revert throwaway edits before baking. Keep install-skip markers or caches outside the workspace directory (e.g. directly under C:\Users\Public\remote-desktop\) — sync mirrors with robocopy /MIR, which deletes anything in the workspace not present locally.

Gotchas

• Daytona shell/process commands run as SYSTEM. GUI apps launched from shell can run invisibly. launch exists precisely for this: it wraps the command in a .cmd and runs it via schtasks /IT /RU Administrator, which puts the window in the interactive session that the noVNC link shows. Always use launch (not shell) to start GUI apps.
• Verify with --wait-window or windows + screenshot — a running process is not a visible window.
• For the shareable link, send desktopUrl (the interactive desktop), not a raw app-port URL.
• record runs Playwright locally: it resolves playwright/playwright-core from --project-path first, else auto-installs playwright-core into skill state, and drives the local Chrome binary (--chrome-path, default macOS Chrome). --mp4 needs local ffmpeg.
• Maintainer note: the launch .cmd generation uses String.raw — keep it that way. A plain template literal turns Windows path backslashes (\node_modules, \bin) into JS escapes (\n, \b) and silently corrupts the launcher.
• Native schtasks /Delete on a missing task writes to stderr, which PowerShell's $ErrorActionPreference='Stop' escalates — the script routes it through cmd.exe /c to swallow that; do the same for any other best-effort native command you add.