AI Harness Engineer (OpenClaw/Hermes) - stabilise & harden agent system

We run a self-hosted, multi-agent AI harness (OpenClaw + Hermes-based control layer) for a small consumer-health company. Today on headless Mac minis, mid-term on Linux VPS; all hosts and team devices connected via Tailscale. It talks to the team over messaging, reads mail, Notion, Drive and meeting transcripts, and runs scheduled briefs and SOPs. Architecture is designed and documented; parts are built, parts aren't, and what's running needs hardening.

YOUR JOB: stabilise what exists, then build out:
- multi-user gateway (shared/private sessions, roles, member on/offboarding)
- memory & shared knowledge base with per-member access control
- skills (versioned, vetted, MCP tool integration)
- cost metering per task/member, model routing cloud/local, enforced caps
- encrypted backups with tested restores, scoped credentials, audit log, emergency stop
- multi-device management (Tailscale ACLs, device onboarding/offboarding, key rotation, nothing on the public internet)
- migration Mac mini -> VPS (containers, hardening, secrets, monitoring) without downtime

MUST-HAVES
- Deep hands-on OpenClaw knowledge (gateway, channels, cron/heartbeats, provider config, rules files, skills); Hermes a strong plus
- You have operated a private network of servers + team devices for a business (Tailscale, Headscale, NetBird, ZeroTier, WireGuard or similar)
- DevOps on macOS (launchd, logging, Ollama, git-based config) and Linux VPS
- Eval mindset: golden test sets, runbooks for backup/restore and incidents
- Very high responsiveness: same-day reaction to incidents, proactive status updates
- Execution focus: small reviewable PRs, no re-architecting of decided plans
- At least one verifiable business reference for production infrastructure or agent/automation work
- Signed NDA + data-processing agreement (plus EU Standard Contractual Clauses if you are outside the EU/EEA) before any access. Work only on our infrastructure; no data leaves our systems.
- Good written English; EU time-zone overlap required

HOW WE WORK
Async, one short weekly sync, Kanban tickets with acceptance tests, all changes via git PRs merged by us. No autonomous changes to production config or credentials; no new tools/integrations without sign-off.

PROCESS
1. Written screening answer (below)
2. One reference call
3. NDA + DPA
4. Paid trial (~1 week, fixed scope): backup + tested restore and a health-check heartbeat with escalation
5. Then ~2-3 months stabilisation, reducing to ongoing hours (retainer)

TO APPLY - bids without this are ignored:
- Link relevant OpenClaw/Hermes work, repos or write-ups
- Name one business reference we may contact
- Answer in under 150 words: "A self-hosted agent silently stopped sending its scheduled morning brief. What do you check, in what order, and what do you tell the owner and when?"