wiki-vault/Sources/Homelab/2026-05-18-claude-desktop-subprocess-leak-memory-management.md

created	path	project	tags	type
2026-05-18	Sources/Homelab	herbys-dev-setup	claude-code dispatch automation	session-notes

Outcome

Identified that Claude Desktop on us-test-authy accumulates Dispatch subprocess zombies that consume substantial RAM (~10.6GB across 41 processes observed during this session). Closed Desktop to recover ~7GB immediately. Parked the long-term fix as a follow-on — daily nightly stop via user systemd timer is the leading approach.

Topics Covered

• Diagnosed RAM pressure on us-test-authy (13/23GB used, hunting for the cause)
• Found 41 claude subprocesses totaling 10.6GB, all parented to the live Claude Desktop electron process (PID 183216) — not orphaned, just never reaped
• Subprocesses ranged from ~6h to 8 days old; oldest were Dispatch sessions long abandoned
• Confirmed the pattern: Desktop spawns claude --output-format stream-json subprocesses per Dispatch session, doesn't SIGTERM them on session end, they sit idle attached to the parent
• Closing Claude Desktop killed the parent and all children together — clean recovery
• Discussed reaper-cron vs. parent-restart approaches; parent-restart wins on simplicity

Key Learnings

• The subprocess leak is structural, not a cron-able janitor problem. The parent is alive and "owns" the children; they don't get cleaned up until the parent dies.
• This is the same failure mode the dispatch-replacement work is designed to eliminate (harness owns subprocess lifecycle explicitly). Tonight produced concrete evidence for why that work matters.
• User systemd timer is the right mechanism, not root cron — Claude Desktop is a GUI Electron app tied to the user session and DBus.
• "Just stop, don't restart" is cleaner — at 4am there's no graphical session to relaunch into, just stop and let user relaunch on next sign-in.

Follow-ons

• Decide whether to implement the user systemd timer (stop-only at 04:00) or rely on manual daily quit-and-relaunch
• If implementing the timer: confirm hour doesn't conflict with any Lovebug overnight work, then write ~/.config/systemd/user/claude-desktop-nightly-stop.{service,timer}
• Verify Manjaro tower exhibits the same leak pattern; if so, deploy same fix via Ansible
• File observation as input to dispatch-replacement design — subprocess lifecycle ownership is a hard requirement, validated by this incident
• Optional safety-net: memory-pressure alert cron (notify if free -h available drops below 4GB)