Fix a “bricked” Grok Bot computer after a RAM OOM (and a slow Reset)

A heavy job on the Agent Computer (for example embedding / graph indexing) can hit the VM’s RAM limit. When memory runs out, the job and system services crash together, so the machine looks bricked and Reset looks broken. Staff on Grokbot bricked, won’t reset confirmed that case: free disk was fine; memory was the failure. Auto-resume on a fresh computer repeated the crash until the workload stayed smaller.

What staff saw

• Reset did work — it was just very slow (close to an hour that day).
• Clicking Reset again while the previous Reset is still running restarts the timer from the beginning.
• Durable working files (about 12 GB on that report) restored after Reset. Packages installed with apt or pip outside those files need a reinstall.
• Updating the desktop app to a current build (0.43.0 on that thread) is the right app-side move; it does not by itself stop an OOM loop.

This is different from a partial-state Reset hang (“Reset failed… may be in a partial state”): Fix Grok Bot after a failed Reset leaves a partial state.

Stabilize after the computer comes back

1. When the computer is up, do not start the heavy indexing job immediately. If a routine or the Bot resumes it on its own, stop it first.
2. In the Bot shell, check memory with free -h before anything large.
3. Run embedding / graph work in smaller batches: fewer documents per run, fewer parallel workers, or a smaller embedding model, so total RAM stays well under the limit.
4. If you need another Reset, click it once and wait up to about an hour before clicking again or starting heavy jobs.

Official order for unreachable computers still applies: Retry → restart app → Recover / Update Agent Computer before Reset. See Troubleshooting.

Pitfalls

• Re-clicking Reset because the UI looks stuck can lengthen the wait. It does not speed recovery.
• “Bot failed to respond” after OOM may need the failed-to-respond path once the computer is healthy again: Understand Grok Bot usage and “failed to respond”.
• Collect version, OS, approximate time and timezone, and whether free -h showed pressure before filing a new report.