Codex Computer Use Makes replayd Spike? Contain It on Mac

If replayd stays near 100% CPU while Codex or ChatGPT Computer Use is idle, stop the active Computer Use task, save your work, and fully quit the desktop app. Wait a minute and check Activity Monitor again. If both replayd and SkyComputerUseService settle, update the app and macOS before retesting one disposable Computer Use task. If the load returns while idle, disable or uninstall Computer Use through the app's supported settings or plugin UI and leave it off while the upstream issue is investigated.

Do not delete replayd, remove Computer Use runtime folders, edit notification hooks, or repeatedly kill system processes. replayd is part of Apple's screen-recording stack. A manual process stop may interrupt the current loop, but macOS or the client can start it again without fixing the trigger.

Confirm the Exact Process Pair

Open Activity Monitor, select the CPU tab, and search for these names:

• replayd

• SkyComputerUseService

• Codex or ChatGPT

You can also take a read-only terminal snapshot:

ps -axo pid=,ppid=,%cpu=,etime=,comm= |
  egrep 'replayd|SkyComputerUseService|Codex|ChatGPT'

Record the app version, Computer Use service version, macOS version, CPU values, and whether a Computer Use or Record & Replay task is active. Check the versions in the app and plugin details rather than inferring them from a folder name.

The strongest current report observed replayd at about 91.2% CPU and SkyComputerUseService at about 19.8% while Computer Use was not active. Its system log counted 2,067 accepted connections from each client and 4,134 cancellations in one second. That is unusually specific evidence for a connection loop, but it is still one environment: ChatGPT desktop 26.730.61639, Computer Use service 26.803.1000621, and macOS 14.6.

A commenter using macOS 26.5.2 and Computer Use service 26.804.1000633 reported no reproduction with a similar hook and runtime. Do not treat the hook, the runtime directory, or replayd alone as a universal root cause.

Prove the Load Follows Computer Use

Use a simple stop-and-observe test before changing configuration:

1. Stop any Computer Use, browser-control, Record & Replay, or desktop-inspection task.

2. Save repository and document changes.

3. Quit Codex or ChatGPT with Command-Q. Closing its last window is not the same as quitting the app.

4. Wait 60 seconds without opening another screen-recording application.

5. Check replayd and SkyComputerUseService again.

Expected result: SkyComputerUseService exits and replayd returns close to idle. If that happens repeatedly, you have a useful correlation with the Computer Use client.

If replayd remains hot after Codex and ChatGPT are fully closed, do not force the diagnosis. Other screen-capture clients can use the same Apple service. Quit active recording, meeting, streaming, remote-desktop, screenshot, and browser-automation tools one at a time. A separate Codex report tied high replayd CPU to leftover headless Playwright or Chrome workers, while another tied it to Chronicle recording. The process name identifies the Apple service, not the client that is keeping it busy.

Contain the Loop Safely

When the load follows Computer Use, use the narrowest supported containment:

1. Leave Computer Use tasks stopped.

2. Update ChatGPT or Codex through its normal updater.

3. Update macOS through System Settings > General > Software Update when your device supports the current release.

4. Reopen the app without restoring a long-running Computer Use task.

5. If the app exposes a Computer Use toggle or installed-plugin control, turn Computer Use off there.

6. Quit and reopen the app, then confirm the helper does not return while idle.

The UI can differ by build and account. Use the visible supported control rather than modifying $CODEX_HOME, plugin caches, app bundles, global configuration, or notification hooks by hand. The August 7 reporter removed a hook and runtime while diagnosing the bug, but that destructive cleanup is evidence from the reproduction, not a general recovery procedure.

If the app is unresponsive, use Apple's Force Quit window after preserving work. Restarting the Mac is a reasonable final containment step when a screen-capture daemon remains hot after all likely clients are closed. A restart clears the current process state; it does not prove the issue is fixed.

Why Killing replayd Alone Is Not a Fix

Apple documents ReplayKit as the framework for recording or streaming screen and app audio. replayd is a macOS service used by that capture path. A client that reconnects immediately can make a replacement replayd process hot again.

That matches the current report: stopping the processes manually did not hold, because the Computer Use service later restarted. The durable diagnostic is whether the owning client remains stopped and whether the load stays down after a clean app restart.

Avoid these tempting responses:

• Do not delete or rename system daemons.

• Do not disable macOS security protections.

• Do not revoke every privacy permission before proving which client is active.

• Do not run broad killall or pkill -f commands against Codex, ChatGPT, Node, Chrome, or recording tools.

• Do not remove the entire Computer Use directory or plugin cache.

• Do not paste raw unified logs into an issue without checking for window titles, paths, account names, or private content.

Broad cleanup can terminate unrelated work and erase the version evidence needed to diagnose a regression.

Separate Similar High-CPU Failures

Several macOS performance failures look alike in Activity Monitor but need different containment.

This classifier matters because quitting Codex can temporarily improve several unrelated bugs. The process pair, version pair, and controlled negative test make the report actionable.

Retest After an Update

Do not restore an unattended desktop workflow as your first test. Use a disposable application and a read-only action:

1. Record the new ChatGPT or Codex, Computer Use, and macOS versions.

2. Confirm replayd is close to idle before starting.

3. Enable Computer Use through the supported UI only if you deliberately disabled it.

4. Ask Computer Use to inspect one disposable window without clicking or typing.

5. End the task and wait five minutes.

6. Confirm SkyComputerUseService and replayd return close to idle.

7. Quit and reopen the app once, then repeat the same test.

Success means the read-only inspection completes, the task ends cleanly, and CPU returns to baseline after both runs. If the loop returns, disable Computer Use again and add your sanitized version matrix to issue 37420 instead of opening a duplicate.

Monitor the Upstream Reports Without Posting

Aident Loadout can check the connected GitHub issue surface without placing a GitHub token in your prompt. Give Codex the canonical setup instruction:

Follow https://aident.ai/SETUP.md

Then use this bounded prompt:

Check my Aident Loadout account and Vault status. Find a connected read-only GitHub issue search Action, inspect its current schema, and preflight it. Return the status, updated time, comment count, affected versions, and any maintainer-authored workaround for openai/codex issues 37420, 30503, 20313, and 28352. Do not create or edit issues, post comments, change files, install software, or ask for a provider key.

Success is measurable: you receive four current issue records, the Action performs zero GitHub writes, and no credential appears in the prompt or result. Set up Aident Loadout and run the read-only check.

Sources

• OpenAI Codex issue 37420: Computer Use and replayd XPC reconnect loop

• OpenAI Codex issue 30503: idle Computer Use helper keeps ReplayKit active

• OpenAI Codex issue 20313: Chronicle drives replayd CPU and capture failures

• OpenAI Codex issue 28352: leftover browser workers keep replayd active

• Apple Developer Documentation: ReplayKit

• OpenAI: ChatGPT Work and Codex

• OpenAI: Use your computer with ChatGPT and Codex

Refresh this guide when issue 37420 changes state, OpenAI publishes a supported recovery, a Computer Use release changes the affected version matrix, macOS changes the ReplayKit behavior, or the linked Aident setup and diagnostic guides change.