Thirteen Tools, Thirteen Releases: Every Greyforge Open-Source Tool, Improved

CRIS set the bar

CRIS is the system monitor we wrote for the Omarchy bar in September. It is a C program with no libc that makes raw syscalls, sleeps between ticks and holds 16 KB of memory. Next to it, btop uses 32 MB.

The question for this sweep was simple: what would each of our other tools look like held to that standard? Less work per call, fewer processes started, fewer allocations, less code where the same job can be done with less, and every claim backed by a number measured on the same machine before and after.

CRIS went through the sweep too. It came out with four sampler fixes and a widget that restarts a dead sampler, and the binary is still exactly 5,512 bytes.

How each tool went through it

Each tool got its own fresh clone and the same written brief. Measure first. Fix bugs and cut work. Add a test for every fix that fails on the previous release. Measure again with the same method, interleaving old and new builds, because the host was busy and medians beat single runs. Then write it all down: numbers, method, what was left alone and why.

An independent review then read every diff and re-ran every suite. That pass earned its keep. It moved Grabbar's drag cancel out of a decoration destructor and into the plugin's exit hook, so nothing calls back into Hyprland while the compositor is tearing a window down. It had slake re-raise a stop signal after recording the run, so systemctl stop reads as a clean stop. It kept typing.get_type_hints() working on memory-quality-gate's public dataclasses.

The four Omarchy plugins were then loaded together in a nested Hyprland compositor on a headless output, with a sandboxed Omarchy shell and its own fake home directory. Grabbar's native plugin passed 19 startup checks and 16 interaction scenarios there. CRIS restarted its sampler 5.2 seconds after we killed it. Nothing ran on the live desktop.

Last came the release gate: a leak scan of every new line and commit message, CI green on the exact commit, then the tag and the release notes.

The cost was in the processes

The same finding kept turning up in different languages: the expensive part of a small tool is usually the processes it starts, not the code it runs.

• firepot, a single bash script, forked subshells for every JSON escape, threshold and percentage. Doing that arithmetic in bash and reading /proc and /sys directly took a full run from 559 processes to 12 and from 886 ms to 76 ms.
• Sley Tools ran its whole doctor check, nine helper processes, before every run, test and edit. It now checks the core by path, pinned SHA-256 and manifest, which took run from 26.1 ms to 11.0 ms.
• devcap started three helper threads per probe. One poll(2) loop over the pipes and a Linux pidfd replaced them: 16 threads per scan instead of 172, and half the CPU.
• Reprieve's Super+W path dropped from four processes to two, Hotbar's hotbar pins from four to one, and oremap's 64-repository scan from 193 to 129 by asking git status for the branch and the changes in one call.

Parsing was the second theme. tongs read and parsed a store twice per CLI command and now reuses the first parse when the bytes under the lock are identical. slake ran seven statements and two table scans every time it opened its ledger; it now records its schema revision and runs one.

Dependencies were the third. slake swapped serde, serde_json and time for small built-in formatters with identical output and built its bundled SQLite without the features it never uses. Its runtime dependency tree went from 45 crates to 24 and its binary from 2.91 MB to 2.13 MB.

Omarchy plugins

The desktop tools got the most correctness work, because they track windows that Hyprland is free to move, close and renumber. Reprieve now drops a pending move when a window closes mid-flight, so a recycled address can never hide a new window, and redo returns a window to where it actually was rather than where it was first parked. Grabbar brings back every window on its hidden workspace when it unloads, even ones it has no record of. Hotbar maps a pin typed in the wrong case to the real app and treats names like constructor as ordinary keys.

All four install the same way, from GitHub through Omarchy's plugin manager. CRIS, for example: omarchy plugin add https://github.com/GreyforgeLabs/omarchy-cris --enable.

Data you can trust

Three of the tools exist to keep data intact, and each got stronger guarantees. moldbox now refuses to publish a backup under a name SQLite would treat as the source's own -wal, -shm or -journal file, and it hashes the synced file before publishing it. tongs sets a new file's mode through the open descriptor and fsyncs the parent of every directory it creates. oremap, which reads env files only to list variable names, now keeps every line of a multi-line value out of that list, PEM blocks included.

Every release card

Each card shows the release and the measured wins. Select one to open its release notes on GitHub.