---
name: migrate-offlab-pending-2026-06-13
description: "Fleet durable state isn't fully backed up to external remotes; Cameron asked to plan a migration to an external host that doesn't depend on the lab server's uptime"
metadata: 
  node_type: memory
  type: project
  originSessionId: cd07777c-8add-4afc-9cb9-9113caa0331f
---

Cameron flagged on 2026-06-13 that the lab server (phe108) has no uptime guarantee — could be shut down or wiped on short notice — and asked to put "migrate the agent codebase + setup to an external host" on the fleet-infra checklist. Already filed in `vault/fleet/infra/overview.md` under "Open infra work."

**Current backup posture (audited 2026-06-13):**
- `agents_stuff` (82M): ✅ in `github.com/cameronosmith/agents_stuff` via SSH remote, pushed yesterday.
- `vault` (2.2M): ⚠️ local git repo exists (commit `0ea4fd0`), **no remote set** — never pushed externally.
- `memory` (44K, manager's persistent memory dir under `~/.claude/projects/.../memory/`): ❌ **not a git repo at all** — zero backup, biggest single exposure.
- `para` repo (where `serve.py` lives): ✅ in github but the remote URL has a PAT in plaintext — same leak pattern as the 2026-06-04 PAT, should rotate to SSH.

**Why:** Disk-only state is one cleanup away from gone. The scaffolding-for-reference README I pushed 2026-06-12 (cameronosmith/agents_scaffolding_for_reference) is intentionally NOT a backup — it's a sanitized public doc.

**How to apply:**
- Don't auto-execute the migration; Cameron wants it scoped and on the checklist, not done immediately.
- When he says "let's do the migration" or similar, the right order is: (1) memory → git + push, (2) vault → add remote + push, (3) daily push cron for both, (4) external VPS that rehydrates from the three repos, (5) test `bootstrap.sh` end-to-end on a fresh VM.
- Hardware-tied agents (yams, backbones, etc.) can't fully relocate — their compute lives on the lab box. The external host is for the durable code/state + the public surfaces (`omidlab.net` dashboard, mobile_chat, voice_interface, our_wandb viewer).
- External hosting options we shortlisted on 2026-06-09 during the wandb/trackio discussion: Hetzner, DigitalOcean, Render, HF Space. Same shortlist applies.

**Tripwire:** if Cameron mentions the lab box being scheduled for downtime / a wipe, surface this checklist proactively — it becomes urgent then.