computer-use: Screenshot calls appear to leak X11 client connections until Xvfb hits Maximum number of clients.

[homanp]

Description

Daytona Computer Use appears to leak X11 clients during screenshot usage.

We start Computer Use successfully and initial health checks pass:

• /computeruse/status
• /computeruse/display/info
• /computeruse/display/windows
• /computeruse/a11y/tree
• /computeruse/screenshot

After repeated agent verification steps that call /computeruse/screenshot?showCursor=true, the screenshot API begins failing with X11 client exhaustion symptoms, e.g. X protocol errors like “Maximum number of clients reached” and/or Toolbox responses like “connection is shut down”.

This looks like screenshot/display/a11y calls may open X11 client connections and not close them reliably, or close them too slowly under repeated calls. Once Xvfb hits its max client count, further screenshot capture is broken for the sandbox even though the Computer Use processes still appear running.

Expected:

Repeated screenshot calls should not monotonically increase X11 client count or eventually exhaust Xvfb clients.

Steps to reproduce:

1. Start a Daytona sandbox with Computer Use.
2. Start Computer Use.
3. Repeatedly call:
   GET /computeruse/screenshot?showCursor=true
4. Observe Xvfb/X11 client count and open FDs for the Computer Use process.
5. Eventually screenshot calls fail with X client exhaustion / connection shutdown.

Impact:

Agents need screenshots for visual verification. Once the screenshot endpoint exhausts X clients, the whole desktop verification flow becomes unusable and subsequent findings are blocked.

[Tushar-Khandelwal-2004]

I opened a PR for this: #5049

I was able to validate the fix with a Docker/Xvfb repro that runs 100 consecutive TakeScreenshot(showCursor=true) calls. The loop completed successfully, and xlsclients showed no lingering X clients afterward.

The fix serializes the X11-heavy Computer Use screenshot/display/window paths and caches primary display bounds to reduce repeated X11 client churn.

[homanp]

I opened a PR for this: #5049

I was able to validate the fix with a Docker/Xvfb repro that runs 100 consecutive TakeScreenshot(showCursor=true) calls. The loop completed successfully, and xlsclients showed no lingering X clients afterward.

The fix serializes the X11-heavy Computer Use screenshot/display/window paths and caches primary display bounds to reduce repeated X11 client churn.

appreciate it!