Finn 128GB Upgrade, RAM Reallocation + Parallel-Agent Unblock

URL: https://mkdocs.justinsforge.com/memory/plans/finn-128gb-unblocks-parallel-agents-2026-06-20/

Date: 2026-06-20. Trigger: Finn (MS-01 Proxmox host) upgraded 32GB -> 128GB DDR5 (125Gi usable, confirmed healthy, ~98GB free at idle).

Why this matters

The parallel multi-agent CreatorTrack dev workflow was hard-blocked on RAM by design. Console (VM 103) was the entire dev budget at 12GB and the design note states it "can't grow without starving production" at 32GB total. The 3-agent trial is built and proven end-to-end but PAUSED at agent-spawn (Console at ~2.1GB free, swap engaged). See parallel-agent-creatortrack-workflow-2026-06-18 section 6b and reference_creatortrack_parallel_dev.

The briefing predicted: "128GB = RAM stops being the limit; bottleneck moves to CPU (Turbopack compile, MS-01 cores; real ceiling ~6-10 actively-building) and merge throughput." The upgrade satisfies the documented resume condition ("bump Console RAM, restart slot servers, /spawn opus one agent per slot").

Reallocation applied (2026-06-20)

Guest	From	To	Method	Status
VM 103 console	12288	51200 (50GB)	qm set 103 -memory 51200	Config set; needs Console reboot
CT 109 forge-data	4096	12288 (12GB)	pct set 109 -memory 12288	Applied live (cgroup re-read showed max; confirm on next CT start)
CT 107 immich	4096	6144 (6GB)	pct set 107 -memory 6144	Applied live

Committed running RAM ~94GB of 125; ~30GB reserved for host + ZFS ARC. No overcommit.

Open follow-ups

1. Console reboot to claim 50GB (drops ALL Console tmux/remote-control sessions; reconnect via /rebridge after). Verify: ssh finn "qm config 103 | grep memory" then ssh console free -g. IN PROGRESS 2026-06-20.
2. Confirm CT 109 cgroup enforcement DONE 2026-06-20: pct reboot 109; host-side /sys/fs/cgroup/lxc/109/memory.max = 12884901888 (12GB), enforced.
3. Retune Postgres on CT 109 DONE 2026-06-20: live values shared_buffers=3.75GB, effective_cache_size=11.25GB, work_mem=32MB. NOTE: shared_buffers + effective_cache_size are driven by an existing postgresql.auto.conf (ALTER SYSTEM), which shadows the block appended to postgresql.conf. Values are correct for 12GB; reconcile the dual source later so the conf block isn't misleading.
4. Resume the 3-agent trial per section 6b: restart slot dev servers (forge_suite_agent_bootstrap.sh <branch> <port>), then /spawn opus one agent per slot. With 50GB Console, run always-on previews instead of round-robin; watch for CPU (MS-01 cores) becoming the new ceiling.