Drowned God Agent-Brick

Headless agent OS profile + Alienware-mothership cockpit TUI for raz-gpd4.

Strip the laptop down to bare-bones, let the local LLM have all the resources, drive everything from a phone via Telegram/web while the rig itself runs a phosphor-green submersible cockpit on tty1.

What's here

drowned-god/
  systemd/
    agent-brick.target         # custom isolation target (replaces graphical.target)
    llama-qwen-8b.service      # default: Qwen3-8B + Qwen3-1.7B spec-dec on :8080
    llama-qwen36-27b.service   # opt-in: Qwen3.6-27B (no spec dec — SSM wall)
    openclaw-gateway.service   # Telegram + tools + local-llm provider
    drowned-god-tui.service    # cockpit TUI on tty1
  scripts/
    dg-mode                    # switch helper: brick | dev | status
  tui/                         # Rust+ratatui cockpit
    src/main.rs
    src/palette.rs             # phosphor sub-aqua palette
    src/corners.rs             # rotating ASCII corner ornaments
    src/telemetry.rs           # llama / FLM / sysinfo / VRAM polling
  install.sh                   # sudo: install systemd units + dg-mode
  uninstall.sh                 # sudo: revert

Install

# build the cockpit binary
cd ~/projects/drowned-god/tui && cargo build --release

# install the systemd units + dg-mode helper
sudo ~/projects/drowned-god/install.sh

Usage

dg-mode status     # see current mode + service health
dg-mode brick      # isolate to agent-brick (kills GUI session — confirms first)
dg-mode dev        # back to graphical.target

To boot directly into brick mode every time:

sudo systemctl set-default agent-brick.target

To revert:

sudo systemctl set-default graphical.target

What you gain in brick mode

• ~3 GB of system RAM freed (KDE+Plasma+browsers→0)
• iGPU driver still loaded; llama-server runs unchanged
• Network up; SSH up; Tailscale up
• Telegram bot stays online via openclaw → @QwenLo_Bot
• Direct cockpit on tty1; remote access via SSH or Tailscale

What you lose

• No GUI on the device (intentional)
• All debugging via SSH or the cockpit TUI
• Browser-based dev tools require a remote browser pointed at this rig

Web access from anywhere

The rig's :8080 (llama-server, OpenAI-compatible) is reachable on:

• LAN: http://raz-gpd4:8080
• Tailscale: http://100.77.212.27:8080
• Funnel (public): tailscale funnel 8080 then the funnel URL

Drive it from a phone browser, OpenAI client, or the existing Telegram bot.

Cockpit controls (TUI)

• ↑ ↓ — move model selector
• enter — send chat
• esc / q — exit
• More controls land as panes wire up (TODO: tool-call streaming, memory recall, model switch)

Troubleshooting

• TUI crashes: journalctl -u drowned-god-tui -e
• llama-server: journalctl -u llama-qwen-8b -f
• openclaw: journalctl -u openclaw-gateway -f

Aesthetic notes

Phosphor green on abyss, sonar cyan accents, amber warnings, salt-rust idle. Rotating corner anchors at 4 fps (Sierra/LucasArts cockpit feel). Sonar pulse in the footer. All gauges live-update from llama-server /v1/models plus rocm-smi for VRAM and sysinfo for everything else.