GodMode

The context

Juggling several local projects (backends, frontends, workers) and LLM models loaded in RAM turns the command line into friction. GodMode is a unified control panel: a menu-bar icon plus a detailed window. One click to start, stop or restart a service; one shortcut to unload every AI model overheating the machine. Multi-project by design: what works for one project works for all, with no hardcoding.

My role

End-to-end design and development: Rust core architecture, IPC surface, system scanners, LLM runtime detection, React interface and test suite. Built solo.

Key features

• Project dashboard: per-project cards, start/stop/restart per service, a health dot and favorites-then-alphabetical sorting.
• Process lifecycle: graceful SIGTERM shutdown with a grace window then SIGKILL, cleanup of orphaned children.
• LLM control: detection of the main local inference runtimes, with real RAM footprint, model, size, quantization and owning-process attribution.
• Kill all LLMs: a large long-press button, a system-bar item and a global shortcut, with a macOS notification of the number of processes stopped and the RAM freed.
• Memory-pressure panel: detailed system memory (used, wired, compression, swap), a deterministic pressure level, hiding of non-killable daemons.
• Automatic service detection by reading manifests (package.json, Cargo.toml, pyproject.toml, requirements.txt, Makefile), inferring the port and build script.
• Live scan of listening ports (lsof cross-referenced with the current directory) and live log tailing, with a readable crash summary extracted from stderr.
• Developer toolbox: inspecting and freeing a busy port, measuring and purging caches (npm, pnpm, yarn, pip, Homebrew, Hugging Face, Xcode).

Technical challenges

• Cleanly killing a shell-wrapped service: GodMode creates a process group at spawn (setsid) then signals the whole group (killpg, grace window, SIGKILL escalation) and sweeps the remaining children: the cause handled, not the symptom.
• Making user-installed tools visible: an app launched from Finder inherits launchd's minimal PATH; the tool resolves the real PATH once at startup via a login shell, then caches it.
• Attributing an LLM model to its project: walking up the parent-process chain (bounded depth, cycle detection) and matching directories at exact boundaries, with the project match winning over the named owner.
• Identifying the model robustly: a fallback cascade (--model flag, Hugging Face cache, weights file, Python module) with size and quantization extraction, using linear ReDoS-safe regexes.
• Measuring true memory pressure: two-level filtering of system daemons to surface only safe-to-kill processes, and pressure computed deterministically (usage, compression, swap) rather than a plain percentage.
• Scanning continuously without drowning the UI: oversampling before filtering, parallel ps/lsof enrichment, distinct cadences (ports, health) and deduplication.

Scale

A real, tested implementation, not a prototype.

• A native desktop application, genuinely functional and battle-tested, not a mock-up.
• Secure by design: everything runs locally, with no server and no port exposed to the outside.
• A carefully tested core, for reliable behaviour on every launch.

Results

An active tool, used daily to orchestrate a multi-project development environment.

• A tool in daily use: starting, stopping and supervising every local project in a single gesture.
• A modular Rust architecture (registry, process manager, scanners) that drives any repository without hardcoding.