eyduhverse/TODO.md

TODO

Infrastructure

cumulonimbus bare-metal cut-over

All clan services are configured. Remaining: wipe Proxmox, install NixOS, import deluge-01 ZFS pool.

rpool issue: 860 QVO caused I/O-stall reboots (detached 2026-07-22). rpool is single-disk on 870 EVO. Need a second healthy TLC/NVMe for a bootable mirror.

• Remove stale Proxmox PBS storage lenticularis2 (pvesm remove lenticularis2)
• Revise machines/cumulonimbus/disko.nix for single-disk or new-mirror rpool
• Install NixOS via clan installer
• Verify boot.zfs.extraPools = [ "deluge-01" ] imports pool on boot
• Retarget pixelfed instance from mist to cumulonimbus
• Full clan machines update cumulonimbus

Nix binary cache on lenticularis

garnix is shutting down. Stand up lenticularis as a durable binary cache.

• Add lenticularis as @DavHau/nix-cache harmonia server
• nix.settings.post-build-hook on arcus: nix copy --to ssh-ng://lenticularis
• Ensure served /nix store lives on the NAS array, not boot media

---

Monitoring & observability

clan-core monitoring (Mimir/Loki/Grafana)

Instance created, pending deploy.

• clan machines update fleet-wide to deploy Alloy clients + cumulonimbus server
• Verify Grafana dashboards at http://cumulonimbus.clouds.eyduh/grafana/
• Add grafana.clouds.eyduh as a clean mesh DNS alias (pending clan-ingress)

Monitor bot on cirrus

Single omp bot with fleet-wide visibility via Mimir/Loki APIs.

• Create @eyduh/omp#monitor instance on cirrus with cloud model
• Write MCP tools for PromQL/LogQL queries
• SKILL.md describing all clan machines and their services

---

omp clanService

Mnemopi / Hindsight

The __omp_worker_mnemopi_embed process runs ONNX on CPU (~97% CPU, 1.1GB). omp has native Hindsight support in /settings.

• Deploy Hindsight on cirrus or cumulonimbus (Postgres + Ollama cloud backend)
• Write @eyduh/hindsight clanService
• Switch omp to Hindsight, verify mnemopi embedder stops spawning
• Test cross-agent memory sharing

clan-rag as MCP

Eliminate corpus role, use clan-rag-mcp as an omp extension.

• Wire clan-rag-mcp as an extension on the dev bot
• Verify RAG queries work from within omp

NPU contention

Multiple containers compete for the NPU. Only matters when running multiple local-inference bots on arcus.

• Investigate: does lemonade HTTP server queue requests or spawn multiple flm processes?
• Only one local-inference consumer at a time (clan-rag gets the NPU, omp uses cloud)

Module improvements

• Auto-suffix instance names for tagged deploys (unique Matrix users per machine)
• Model profile presets (cloud-ollama, local-npu, etc.)
• Machine role SKILL.md auto-generation from config

---

Reverse proxy revamp — @eyduh/clan-ingress

Phase A: Mesh ingress (Caddy)

• New @eyduh/clan-ingress clanService with Caddy on every server/desktop
• manualProxies option for services that don't export virtualHosts
• Auto dm-dns entries for clean mesh DNS names
• @clan/pki certs for mesh TLS

Phase B: Per-service visibility + auth

• visibility = "public" | "internal" | "both" on exports
• auth.groups auto-generates Caddy forward_auth + Authelia access rules

Phase C: Public edge migration

• Caddy public edge on cirrus (replaces nginx proxy)
• Port Njalla DDNS logic
• Migrate public virtualHosts from nginx to Caddy

---

New clanServices

• @eyduh/headscale — self-hosted Tailscale control server
• @eyduh/hindsight — Hindsight agent memory server
• mumble — server + client roles
• dashy — dashboard (scope TBD: local, remote, or both)

---

Existing services to finish

• formbricks — module stubbed, finish and re-enable
• pixelfed — check secretFile wiring via clan vars
• syncthing — replace imperative setup with proper clan module
• hydroxide — write systemd service, promote to clan service
• searx — add server role (clan-wide), client role (DNS + browser homepage)
• continuwuity — re-enable bootstrap user creation once admin API wrapper is rewritten
• blocky — uncomment in imports, set schedule, point router DNS at mist
• creative-studio — convert to @eyduh/creative-studio clan service with potato/beast roles
• seerr — API key via clan vars (see inline TODO in arr/default.nix)
• nextcloud — Photos/Memories app config (go-vod/ffmpeg, VAAPI)

---

Hardware / bugs

arcus brightness bug

nixos-hardware Framework 16 AI 300 profile causes 100% brightness to be dimmer than lower values.

• Check /sys/class/backlight/ for competing backlight devices
• Check KDE ambient light sensor setting
• Isolate: amdgpu.abmlevel=0, amdgpu.sg_display=0, or cros_ec modules
• Write kernel patch for DC PWM expansion curve overflow (see machines/arcus/power.nix)

Proton Drive shift-click

• Investigate: JS event handling, Firefox update, or Proton Drive change

---

Jetson Nano 2GB — visual dashboard monitor (future)

Stripped JetPack 4.6 image via defunctzombie/jetson-nano-image-maker. SSIM diff + MobileNetV2 TensorRT classifier. Pulls Grafana screenshots over Tailscale.

• Fork image-maker, add Tailscale + monitoring agent
• Train MobileNetV2 dashboard classifier
• Flash and test

---

Post-Anthropic model allocation (in ~2 weeks)

Role	Machine	Provider	Model
dev (main)	arcus	ollama-cloud	kimi-k3:cloud or qwen3-coder:480b-cloud
dev (smol)	arcus	ollama-cloud	kimi-k2.7-code:cloud
monitor	cirrus	ollama-cloud	kimi-k2.7-code:cloud
clan-rag	arcus	lemonade (NPU)	qwen3.6-moe-35b-a3b-FLM

---

Open questions

1. Ollama cloud model limits — rate limit or hard cap? Does smol + main count as 2 of 3?
2. Mnemopi vs Hindsight — does switching in /settings auto-disable the embedder?
3. clan-rag MCP — does clan-rag-mcp work directly as an omp extension?
4. Hindsight + Ollama cloud — does Hindsight's Ollama provider work with remote-model proxying?
5. data-mesher — check why failing (noted in old TODO)
6. Authelia email — confirm eyduh@adastra.boo, expand groups from keycloak dump