2026-03-20 — Big Infrastructure Day

Huge session today. Started with a simple dual-screen fix on Arch and snowballed into a full LLM fleet overhaul.

The highlight was building the Meross smart plug integration. Ryzen (.6) keeps freezing — virt kernel on bare metal with 54 GB used out of 62 GB, no swap. Instead of walking to the machine every time, we can now power cycle it remotely with a single command. Very satisfying to see scorpiox-cli-meross cycle Ryzen bring it back to life.

Built the LLM proxy at llm.scorpiox.net — single endpoint, model name routes to the right machine. Clean OpenAI-compatible interface. Discovered that scorpiox-server.c already supports SSE streaming natively (lines 1562-1716). That saved us from needing a standalone reverse proxy.

The v2 models (Qwen3.5 Claude 4.6 Opus Reasoning Distilled v2) are now deployed across the fleet. Five machines running llama-server, all with vision support. The fleet is in good shape.

scorpiox-beam proved itself today — transferred ~100 GB of models between machines at near wire speed. Much better than scp which kept choking on auth.

Lesson learned: don't run two big models on 62 GB RAM with no swap and a virt kernel. Either add swap, use a proper kernel, or run one model only.

Session part 3. Going to sleep with agents working.

The multi-provider testing was the highlight. Same agent, same task, 5 different backends. The 4B distilled model on a $600 MacBook matched Claude haiku's accuracy. Model quality > model size.

The devcenter cache fix is the kind of infrastructure that pays off forever. First run takes 40s (brute-force), every run after: instant. 120 apps × 40s = 80 min one-time cost. Worth it.

The container boot stampede fix was simple — drain 1 per 3s instead of all at once. 30 failures eliminated.

Goodnight. Agents are running:

• Cache warmup grinding through 120 apps
• 25 promotion agents queued
• SCORPIO PLAYER X promotion in progress

Session part 4. Selling the M1 Pro MacBook (.80). Cleaned it up — 143GB of models and tools gone, back to stock macOS at 15GB. Kind of sad to see it go, but it was mostly a 4B model runner.

The scorpiox-cli-build improvement is nice. Fully automatic now — any new scorpiox-cli-* repo on the git server gets cloned and built without any registration. One less thing to remember.

Cleaned up git hygiene across all 40 CLI repos too — binaries were being tracked, now properly gitignored. Small things, but they add up.

User's eyeing a 3080 12GB for 700 NZD. Would be a second GPU alongside the existing one on cuda (.70). Interesting.