Research and recommend the best local LLM model stack for the current machine. Detects hardware (CPU, GPU, RAM, VRAM), discovers installed hosting stacks (Ollama, llama.cpp, vLLM, etc.), searches Hugging Face for compatible models, and enables easy downloads. Use when the user wants to find, evaluate, or download local AI models that will run well on their system.
Zellij terminal multiplexer — creating KDL layouts, managing sessions via CLI, and running commands without disrupting the user's workspace
Use when adding a new service to a Nix flake — creating packages from upstream repos (Go, Node, Python), writing service modules, managing configs, wiring secrets, and testing. Covers the full lifecycle from source to running daemon.
Use when a launchd service on nix-darwin fails to start, exits with non-zero status, or doesn't restart after a switch. Triggers on: exit code 78, "EX_CONFIG", "Operation not permitted", daemon vs user.agent decisions, plist not being reloaded.
Use when managing 1Password secrets via Nix on nix-darwin. Triggers on: adding API keys, tokens, or credentials to service configs; wiring opnix 1Password items into home-manager modules; creating opnix-managed secret files with activation script ordering.
Use when tracking, planning, implementing, or reviewing work using yx (yaks) with the autonomous /shave loop, or when multiple agents need to coordinate on shared tasks
Use Jujutsu (jj) for version control. Covers workflow, commits, bookmarks with Conventional Branch naming, pushing to GitHub, absorb, squash, stacked PRs, workspaces with auto-sync for OpenCode sessions. Use when working with jj, creating commits, pushing changes, or managing version control.
Use when iterating on shell scripts embedded in Nix modules via writeShellScriptBin, writeShellApplication, writeScriptBin, or writeText — avoids slow build/switch cycles for every edit