macOS: Computer Use-correlated FullWake cycles followed by WindowServer/AGX GPU exhaustion

What version of the Codex App are you using (From “About Codex” dialog)?

26.901.20858 (Computer Use helper 26.831.1000926)

What subscription do you have?

ChatGPT individual paid plan

What platform is your computer?

macOS 14.8.9, Apple silicon (device model and unique identifiers intentionally omitted)

What issue are you seeing?

On one overnight interval, the Mac was locked with the lid closed and connected to AC power. It later became unusable: an external display showed only a large solid-color block, while the built-in display showed only a charging indication instead of a usable login/session UI. A forced power-off was required.

Post-incident log reconstruction found a strong temporal correlation between Computer Use activity and repeated wake/display reconfiguration:

• macOS recorded 26 HID-triggered FullWake events during the locked interval.
• 24 of 26 occurred in the same second or within 2–6 seconds of SkyComputerUseService creating a "Codex Computer Use interaction" UserIsActive/display assertion.
• The first display/GPU failure appeared immediately after one such wake: "invalid display identifier" in the ChatGPT/Computer Use path, followed by WindowServer "invalid display seed", Metal kIOGPUCommandBufferCallbackErrorOutOfMemory, and IOGPU "failed to create mapping" errors.
• The failures increased after later wake/display cycles. WindowServer Metal out-of-memory events reached roughly 1,955 in one minute, and the final minute before restart contained about 60,000 IOGPU mapping failures.
• No kernel panic was recorded. The reset record is consistent with a forced power-button shutdown.
• After reboot, the same error classes stopped.

I am reporting a strong temporal correlation, not claiming that Computer Use is already proven to be the sole root cause.

What steps can reproduce the bug?

1. Enable Codex Computer Use and permit locked use.
2. Leave the Mac locked, lid closed, and connected to AC power overnight. No foreground Computer Use task was intentionally left running.
3. During the locked interval, observe repeated HID FullWake events closely following SkyComputerUseService "Codex Computer Use interaction" assertions.
4. After enough wake/display cycles, attempt to resume using an external display and/or open the lid.
5. The display stack may fail to render a usable session and show a solid-color block, requiring a forced shutdown.

I have not deliberately repeated this test because the observed failure can leave the machine unrecoverable without a hard power-off.

What is the expected behavior?

When no active UI task is intentionally running, Computer Use should not repeatedly pull a locked, sleeping Mac into HID FullWake. Repeated locked-use wake/display transitions should not exhaust WindowServer or AGX/IOGPU resources, and opening the lid or attaching a display should always restore a usable login/session UI.

Additional information

Please investigate the interaction between Computer Use, Locked Use, and macOS sleep/wake display reconfiguration—especially whether background Computer Use calls can trigger FullWake and retain or exhaust display/GPU mappings across wake cycles. Computer History was enabled on the affected setup, but its involvement is unproven.

Representative redacted errors:

ChatGPT / SkyComputerUseService: invalid display identifier
WindowServer: PKGWindowMoveOnMatchingDisplayChangedSeed ... invalid seed
Metal: Insufficient Memory (kIOGPUCommandBufferCallbackErrorOutOfMemory)
IOGPUResource::map ... failed to create mapping
AGX driver: Failed to init GTP channel

Related but different: #33471 also reports "invalid display identifier" around Locked Computer Use/display transitions, but its primary symptom is remote takeover failure rather than repeated sleep wake followed by GPU resource exhaustion.

No screenshots or raw logs are attached because they may contain device or account identifiers. I can provide a narrowly filtered, further-redacted excerpt if a maintainer specifies the exact fields needed.

Begin with the Computer Use/Locked Use path represented by SkyComputerUseService and compare the reported sequence with issue #33471 and the filtered macOS logs, focusing on FullWake and display assertions. Done means establishing whether background Computer Use can cause repeated locked-sleep wake/display transitions and preventing the resulting WindowServer/AGX/IOGPU exhaustion while preserving a usable login/session UI.