Blog Setup
Goal: Static site + LLM CLI dev environment without giving any tool shell access to personal Mac.
Stack
- Eleventy (11ty) — markdown-first, minimal config, no magic
- Nunjucks templates
- Plain CSS, no framework
- Hosted on GitHub Pages initially; later self-hosted on k3s
LLM CLI Selection
Codex CLI: Installed, rejected. UX was confusing (semi-transparent prompt UI, constant helper text), slow.
Gemini CLI on Mac: Triggered Home folder access prompt on launch. Rejected — shell-level disk access on a personal machine is too broad a trust grant, regardless of whether it was benign.
Solution: Ubuntu 22.04 Hyper-V VM on Windows host. Gemini CLI installed inside the VM. Disk access is scoped to the VM's filesystem.
Static Site Generator Selection
Tried Hugo, Jekyll, Astro. Landed on Eleventy — markdown files, simple Nunjucks templating, output goes exactly where you tell it. No framework opinions.
GPU Passthrough Exploration (dead end)
Wanted local inference (Ollama + 5070 Ti) inside the VM.
PCIe passthrough: Requires level 1 hypervisor (bare metal). Hyper-V is level 2 (runs under Windows). Not viable without reinstalling the OS.
GPU para-virtualization: Carves a static VRAM partition for the guest. Issues:
- VRAM partition is permanent — capacity lost even when VM is off
- No working Windows→Linux path on 50-series; drivers are reverse-engineered and fragile
- Community hacks pull from WSLg kernel source to get Nvidia support in Linux guests — breaks on any driver update (~2/month from Nvidia)
Outcome: Skipped local inference. Fallback path: run Ollama on Windows host, access from VM via subnet endpoint.