You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
November 21, 2026; all times below are CST (UTC−06:00)
Steps to reproduce
Ask a Dot connected to the Mac to open Calculator with Computer Use, calculate 38 × 45, take a screenshot, and finish.
Leave Calculator open. Don't reset Computer Use, click the sharing control, or start another Computer Use task.
After the Dot turn ends, check the macOS sharing indicator and compare the unified logs for SkyComputerUseService and replayd for several minutes.
Actual behavior
Two ScreenCaptureKit streams started at 07:24:56.996 and 07:25:00.766. The Dot's last Calculator screenshot was at 07:25:01.364, and the turn then ended. I found no matching SCStream stopCaptureWithCompletionHandler:, RPClient stop, stopAndInvalidate, or session deallocation through 07:28:15. At 07:28:11.106 and 07:28:12.332, replayd still reported 273 and 283 screen frames in the two sessions' respective health samples. So this wasn't just a stale indicator.
In an earlier native Codex Calculator run on the same Mac, Sky called stopCapture for both active streams at 07:21:17.943 and 07:21:17.948. replayd logged both stops and both sessions were freed by 07:21:17.951, about 32.6 seconds after that Codex turn ended.
Expected behavior
When a Dot turn finishes, release the capture streams it owns unless another active task needs them. If capture intentionally stays active, show why and give the user a clear way to stop it.
What the logs establish
The first missing event in the Dot trace is the stop request from Sky to ScreenCaptureKit. The logs don't establish whether the Dot's cleanup request was never sent, was lost in the handoff, or reached the helper without stopping the streams. The runs were on the same Mac, but the native Codex run included additional interactions, so the 32.6 seconds is an observation, not a timeout. These logs show continued local capture, not that frames were saved or transmitted after the turn.
Could the team trace Dot and native Codex turn completion through to the helper's stop calls? This is related to #50921, but I'm filing it separately because the logs show frames continuing after the Dot turn, whereas the original report described menu entries remaining after capture had stopped.
Activity