CLAUDE.md

Claude Configuration Notes - Dotfiles

Neovim Autosave Issues

Problem

Neovim was not autosaving on all edits. The configuration used TextChangedPost event which doesn't exist in Neovim.

Solution

Fixed by updating the autocmd to use the correct events:

• TextChanged - triggers after text changes in normal mode
• TextChangedI - triggers after text changes in insert mode
• CmdlineLeave - triggers after leaving command line (catches substitutions and ex commands)

Location: /home/tom/git/dotfiles/nvim/init.lua lines 183-189

Current Autosave Triggers

1. InsertLeave - when leaving insert mode
2. FocusLost - when neovim loses focus
3. TextChanged, TextChangedI, CmdlineLeave - on text changes
4. BufLeave - when leaving a buffer
5. Periodic checktime every 2 seconds to reload external changes

i3 Function Key Bindings

Problem

Function keys F1, F2, F3, F5, F9 were not working in applications because they were bound in i3 config.

Solution

Commented out the function key bindings in i3/config:

• F1-F3: concentration label tracking
• F5: inference signal file
• F9: log timestamp refocused script

Reload i3 with $mod+Shift+r to apply changes.

Semantic Code Search Plan

Goal

Add local semantic search alongside existing ripgrep/fzf flows using a persistent vector index.

Components

1. Neovim plugin: vectorcode.nvim for Telescope pickers and CLI utilities.
2. Embedding runtime: Ollama running nomic-embed-text locally (HTTP API compatible with OpenAI format).
3. Vector database: Qdrant in Docker with storage volume at ~/qdrant_storage.
4. Updater: Watchman + Python helper that batches modified file paths and calls vectorcode index --update.

Implementation Steps

1. Start services
   • docker run -d --name qdrant -p 6333:6333 -p 6334:6334 -v ~/qdrant_storage:/qdrant/storage qdrant/qdrant
   • ollama run nomic-embed-text (ensure Ollama daemon is up).
2. Configure VectorCode
   • ~/.config/vectorcode/config.json → set embedding base URL to Ollama and database to the Qdrant collection.
   • Run vectorcode index --rebuild $HOME/git/dotfiles once to seed the index.
3. Automate updates
   • Watchman trigger on repository BufWritePost / created files writes paths to /tmp/vectorcode_queue.
   • Cron or systemd timer executes vectorcode index --update @/tmp/vectorcode_queue every ~30 s, then clears the queue.
4. Neovim integration
   • Map <leader>sv → :VectorCode search (semantic) and keep <leader>s for live_grep.
   • Optional: add :VectorCode refresh command to Neovim which touches the queue file for manual reindexing.
5. Maintenance
   • Weekly vectorcode index --sync to catch branch switches or large refactors.
   • Monitor ~/.vectorcode/logs for failed ingests and re-run update if needed.

Notes

• Keep chunk size consistent (default 512 tokens) so Qdrant payloads remain uniform.
• If Ollama is too slow, swap in a sentence-transformers model inside a Python FastAPI server that mimics /v1/embeddings.
• For multi-repo support, use separate Qdrant collections and include {project = "dotfiles"} in payload metadata for filtering.

Hyprland Configuration

Setup

• Config: ~/git/dotfiles/hypr/hyprland.conf (symlinked)
• Display: Samsung S95D 55" OLED at 4K@120Hz
• Terminal: Ghostty with frosted glass blur

Keyboard

• ZSA Voyager (compact split, no arrow keys)
• Colemak DH layout
• Navigation: Super + neio (focus), Super + Shift + neio (move window)

HDR + Transparency (Fixed Jan 2026)

Use cm,hdr in monitor line instead of cm_enabled = true in render block. This gives HDR AND working transparency.

monitor=HDMI-A-1,3840x2160@119.88,auto,auto,bitdepth,10,cm,hdr
render { cm_enabled = false }

Style

• Gaps: airy (12 inner, 40 outer)
• Blur: frosted glass (size 6, passes 3)
• Animations: snappy (~2x default speed)

Notifications: swaync (not mako)

Use swaync (SwayNotificationCenter), not mako. Both were enabled causing DBus conflicts.

Fix: pkill mako && systemctl --user restart swaync Permanent: sudo systemctl --global disable mako.service

Ghostty Configuration

• Config: ~/git/dotfiles/ghostty/config (symlinked)
• Font size 7 is normal for 55" 4K (same PPI as 27" 1080p)
• Transparency works with HDR (using cm,hdr in monitor line)

Zsh Ctrl+J Fuzzy Directory Jump

Ctrl+J searches directories starting from ~ (home), not current directory.

Problem

Ctrl+J (fzf_cd) was hanging for several seconds after selecting a directory.

Root Cause

The for loop (for d in {1..20}; do find...; done) | fzf continues after fzf exits. Each remaining find spawns, gets SIGPIPE, exits - but loop runs all 20 iterations.

Solution

Added || break to exit loop on SIGPIPE: find ... 2>/dev/null || break

Location: /home/tom/.zshrc lines 48-56

Note: Same issue exists in fzf-file-widget (Ctrl+A, lines 90-97).

Bambu Studio on ARM64 (DGX Spark)

Problem

Bambu Studio only provides x86_64 builds. No native ARM64.

Failed: box64

WebKit2GTK dependency blocks emulation. Box64 can't handle webkit symbols.

Working: Docker x86 Emulation

Requires: qemu-user-static package.

Container: bambu-studio (Docker, x86_64 Ubuntu 24.04) Launch: bambu-studio or app launcher Files: ~/3d-prints mounted to /root/3d-prints in container

Cleanup (optional)

• ~/VMs/bambu-studio/ - 5.8GB Ubuntu ISO, not needed
• ~/Applications/x86_64-libs/ - box64 libs, not needed

Steam Gaming on DGX Spark (ARM64)

Overview

Canonical built an ARM64 Steam snap using FEX emulator for x86 translation. The DGX Spark is the primary test device for this effort.

Installation

sudo snap install steam --candidate

Requires NVIDIA driver 580.95.05 series.

Performance (Cyberpunk 2077 @ 1080p)

Configuration	FPS
Medium, no DLSS (Box64)	~50
Low settings (Proton)	~100
High + RT Ultra + DLSS 4 + MFG	175+

Optimal Setup

• Use Proton 10.0-2 beta (not Box64)
• Enable DLSS 4 with Multi-Frame Generation
• Per-game: Right-click → Properties → Compatibility → Force Proton

GB10 vs RTX 5090

• Raw performance: GB10 ~1/3 of RTX 5090
• Memory bandwidth: 273 GB/s vs 1,792 GB/s (6.5x slower)
• Advantage: 128GB unified memory (no VRAM limits)

Confirmed Working Games

• Counter-Strike 2: "Multi-hour sessions, smooth"
• Dota 2, Portal 2: Native Linux, no issues
• Cyberpunk 2077: Works with DLSS/Proton setup

Free Games for Testing

• Quake II RTX: Best ray tracing demo, tiny download
• Warframe: Best looking F2P, good Proton support
• Counter-Strike 2: Confirmed working on Spark

References

• https://discourse.ubuntu.com/t/call-for-testing-steam-snap-for-arm64/74719
• https://www.phoronix.com/news/Steam-Snap-ARM64-FEX

Waybar Spark Cluster Monitoring

Custom modules: spark1.sh, spark2.sh, spark3.sh (SSH, shows "offline" if unreachable).

CPU Calculation

Must count iowait ($6) as idle, not just idle ($5):

100-(($5+$6)*100/($2+$3+$4+$5+$6+$7+$8))

Swappiness

High iowait (70-80%) despite free RAM was caused by vm.swappiness=190.

Fix: sudo sysctl vm.swappiness=60

To clear existing swap: sudo swapoff -a && sudo swapon -a

Neovim 0.11 Plugin Updates (Jan 2026)

Problem

After upgrading to Neovim 0.11.5, startup errors appeared:

1. nvim-lspconfig deprecation warning for require('lspconfig')
2. nvim-treesitter failed - module 'nvim-treesitter.configs' not found

Root Cause

Plugins were pinned to old versions for Neovim 0.9 compatibility. New nvim-treesitter completely rewrote its API - no more nvim-treesitter.configs.

Solution

1. Removed version pins from all three plugins
2. Updated treesitter config to new API:
   • require("nvim-treesitter").install({...}) instead of ensure_installed
   • vim.treesitter.start() in FileType autocmd for highlighting
3. Updated lspconfig to use vim.lsp.config and vim.lsp.enable()

New Treesitter API (Neovim 0.11+)

Old API (pre-rewrite):

require("nvim-treesitter.configs").setup({ ensure_installed = {...}, highlight = { enable = true } })

New API:

require("nvim-treesitter").install({ "lua", "python" })
vim.api.nvim_create_autocmd("FileType", {
  pattern = { "lua", "python" },
  callback = function() vim.treesitter.start() end,
})

New LSP API (Neovim 0.11+)

Old: require("lspconfig").pyright.setup({}) New: vim.lsp.config.pyright = {} then vim.lsp.enable({ "pyright" })