Project spawn cgroup migration 2026 05 04
On 2026-05-04 the forge-remote-bridge cgroup-kill problem (spawned tmux+claude sessions dying on bridge restart) was partially fixed:
Done:
- sudo loginctl enable-linger justinwieb (user.slice survives logout)
- KillMode=mixed drop-in at /etc/systemd/system/forge-remote-bridge.service.d/killmode-mixed.conf
- forge_spawn_session.sh wraps tmux new-session with systemd-run --user --scope --collect --unit=forge-spawn-<name>, bare-tmux fallback if user-systemd unreachable
- 11 stale tmux sockets swept
Not finished (as of session end):
1. No end-to-end smoke test (spawn via bridge → restart bridge → confirm survival)
2. /resume skill still uses bare tmux new-session — only worked this session because VS Code shell had $TMUX=homebase in env; bridge-spawned /resume would land in bridge cgroup again
3. reference_remote_session_spawning.md not updated with new cgroup-escape architecture
4. Coordinator-bot service (forge_telegram_inbox_brain.py) unit KillMode not audited
5. memory-cap.conf note not written (bridge cgroup no longer caps spawned claudes — intentional tradeoff per May-3 OOM evidence)
Why: Root cause was the 2026-05-03 OOM/manual reboot driven by accumulated spawn memory inside the bridge cgroup.
How to apply: Next time spawn or resume work comes up, start by running the smoke test and hardening the /resume skill before declaring this fixed.
[auto-memory session b0b9f224-d44d-4259-b67c-1e439fad673d, confidence 0.85, mode direct]