My dotfiles

Arch Linux. zsh + tmux + Neovim, symlinked with GNU stow.

Setup

git clone https://github.com/aura-zero/dotfiles.git ~/dotfiles
cd ~/dotfiles && ./bootstrap.sh

That's the whole thing. It's safe to re-run — every step checks before it acts.

If stow reports conflicts with config files already on the machine:

./bootstrap.sh --adopt   # pulls those files INTO the repo — check `git diff` after

What bootstrap.sh does

Step	Detail
Base tools	git, stow, base-devel
AUR helper	builds yay-bin if yay is missing
Packages	everything in scripts/{pkgs,devTools,sysPkgs}.txt — repo packages via pacman, the rest via yay
Symlinks	stow into $HOME
tmux	clones TPM, installs plugins headlessly (no prefix + I needed)
zsh	clones zinit, bursts the turbo-loaded plugins
Neovim	lazy.nvim self-bootstraps, then a headless Lazy! sync
Login shell	chsh to zsh

Flags: --adopt, --skip-packages, --help.

Left to you afterwards, because they need your credentials or a private backup:

• atuin sync — atuin register / atuin login. History works locally without it.
• paid fonts — restore into /usr/local/share/fonts/personal/.
• reboot, then re-run ./install.sh — on a fresh machine with an NVIDIA card. The driver is installed but nouveau owns the GPU until reboot, so the first run can't pin voxtype to it (it says so, and skips). The second run is seconds.
• log out and back in — for the zsh login shell, and the input group, to take effect.

voxtype

install.sh handles it. These are the ones that bite, and every one of them fails silently -- voxtype status says idle through all of them:

• The daemon starts blind unless it waits for the compositor. systemd reaches default.target at login; Hyprland publishes WAYLAND_DISPLAY into the user manager only afterwards. Started before that, voxtype has no display -- and both output drivers are Wayland clients (wtype, and the wl-copy that fallback_to_clipboard falls back to), so every transcription is typed into nowhere while nothing logs a failure. scripts/voxtype-wait-wayland.sh blocks in ExecStartPre until the display exists. The only visible symptom was the OSD exiting with "Failed to open display", which reads as a cosmetic overlay problem and is not.

• The CLI needs the pin too, and gets it from zsh. The daemon reads the dGPU index from its systemd drop-in, but ~/.config/environment.d/ never reaches an interactive shell -- it is read by systemd --user, and Hyprland isn't a systemd unit here. So voxtype transcribe ran on the iGPU: 29.31s vs 3.95s on the same clip. .config/zsh/voxtype.zsh sources the generated file.

• The OSD is a replacement, not a theme. The default gtk4 frontend hardcodes its blue waveform; its only knobs are size, margin and gain. The quickshell frontend is the styleable one and is broken in voxtype-bin 1.0.0-2: StyleLoader.qml and RecipeRenderer.qml are declared in the package's qmldir and not shipped, so style, palette, plugin_path and [[osd.visual.layers]] are all inert. .config/voxtype/osd-dotmatrix/ is a complete replacement shell.qml needing neither, pointed at by VOXTYPE_OSD_QML_PATH in a generated drop-in. Its board colour tracks ~/.local/state/caelestia/scheme.json, and its -30 dBFS floor is measured from this laptop's mic -- re-measure on other hardware, recipe in the file.

• The GPU pin is generated, never committed. scripts/voxtype-gpu-pin.sh detects which Vulkan device is the dGPU and writes ~/.config/environment.d/ + a systemd drop-in. ggml always uses Vulkan0, which on this laptop is the iGPU — whisper runs there at 31.85s vs 2.54s for the same clip, with every status command still reporting healthy. The index differs per machine, so it can't be a tracked file. Re-run the script after a GPU change.

• large-v3 does not fit a 4 GB card on voxtype ≥ 1.0.0. It loads fine (3.1 GB) and the daemon reports healthy, then every transcription fails with ErrorOutOfDeviceMemory: 1.0.0 allocates two whisper states per utterance. The committed config uses large-v3-turbo, which also won the accuracy benchmark.

• F8 is dead for ~8s after every restart. The hotkey binds only once the model has loaded; until then voxtype status says stopped and the key does nothing, silently. voxtype record start works throughout — so "CLI works, F8 doesn't" is almost always this. Wait for Listening for KEY_F8 in journalctl --user -u voxtype.

Full writeup, including why the config is generated rather than hand-written: ~/Resources/MSB/machine/tools/voxtype/voxtype.md.

What's in here

Path
.config/nvim	Neovim — plugins, LSP, keybindings
.config/tmux	tmux — keybindings, themes, per-host overrides
.config/zsh	zsh modules, sourced by .zshrc
scripts	the package lists bootstrap reads
.config/voxtype	push-to-talk speech-to-text — hold F8, speak, release
.config/starship	prompt
.config/{ghostty,kitty,wezterm}	terminal emulators

Those READMEs are reference — how things are configured and which keys do what. Setup lives here and only here.

Housekeeping

Symlinks created by stow shouldn't be committed:

find . -type l -exec git rm --cached {} \;