Who is eating this potato
⚠️ This project is 100% vibe-coded slop — see Model credits for who to blame; not a single line was written by a human. It has — and will keep having — stupid bugs LLMs can't see. It sends signals to your processes; the buttons do exactly what they say, which is the dangerous part. Use at your own risk.
A quickshell + Hyprland widget: a pill in the bar, a process list on click, and buttons to hold or kill — plus a CLI that tells you what is stalling the machine.
Named after Russian жор — "the gorging".
Two things htop won't do while you are not staring at it:
- CPU is a tick-to-tick delta.
ps -o pcpureports the average over a process's whole lifetime, so a panel that has quietly eaten half a core for twelve hours shows up as 22% while it is actually at 57% right now. - PSI (
/proc/pressure) is sampled, not just load. On a weak laptop a freeze is almost never "something hit 100% CPU" — it is a memory stall: swap is full, the kernel is grinding through reclaim, everything waits. At that moment the CPU ranking names the wrong process, because a stalling machine makes every process look busy. Whenmemory full avg10crosses the threshold, zhor looks for the culprit by footprint and swap instead.
Nothing is ever killed automatically. Deliberately: rustc at 250% CPU is a
correct rustc, and an autokiller would eat exactly that.
git clone https://github.com/Petyok/zhor ~/zhor && ~/zhor/install.shThe installer is idempotent: it symlinks ~/.config/quickshell/zhor and
~/.local/bin/zhor, appends two lines to hyprland.conf (with a backup), and
starts the widget immediately — no relogin.
exec-once = qs -n -p ~/.config/quickshell/zhor
bind = SUPER SHIFT, ESCAPE, exec, qs -p ~/.config/quickshell/zhor ipc call popup toggle
Removal: ~/zhor/uninstall.sh drops the symlinks and cuts its own block back
out of the config, byte for byte.
Requires quickshell and python3. Nothing else — not a single pip dependency.
The pill shows PSI CPU% · free memory · swap used. The dot on the left goes
green / amber / red with pressure and blinks on an alert. Click to open the list.
It hides itself over a fullscreen window.
The list (SUPER+SHIFT+ESC; Esc or a click outside closes it) is the top by
CPU and the top by RSS merged, plus swap and hot — how many seconds a process
has been over the line. Clicking the CPU% / RSS headers sorts.
| button | what it sends |
|---|---|
⏸ / ▶ |
SIGSTOP / SIGCONT — freeze a process without losing its state |
↓19 |
renice -n 19 — let everything else have the CPU first |
✕ → KILL |
first click SIGTERM, second one (within 5s) SIGKILL |
Buttons are greyed out for processes you don't own: without root kill returns
EPERM, and a button that silently does nothing is worse than a disabled one.
Alerts go to notify-send, once per process while the condition holds:
sustained— a process over the line for a minute (≥90% of a core or ≥25% RAM);stall— the machine is genuinely blocked on memory (memory full avg10≥ 10%).
install.sh puts a symlink in ~/.local/bin/zhor, so:
$ zhor
VERDICT MEMORY TIGHT: 1391M free, swap 100% full, biggest footprint Telegram(1079) — a stall is one heavy tab away
PSI cpu 2.3% mem 0.0% io 0.0% (avg10 = share of the last 10s spent stalled)
MEM 1.4G free of 7.7G swap 100% used
TOP CPU gjs(2232349) 59% claude(1032098) 12% claude(1588326) 5%
TOP MEM Telegram(1079) 344M+439Msw claude(3239768) 483M+107Msw
RECENT 3 alert(s) logged: 2x sustained gjs, 1x stall Telegram
$ echo $?
1| command | what it does |
|---|---|
zhor |
one-shot snapshot and verdict, exit 1 if the box is under pressure |
zhor --json |
the same, machine-readable, with verdict and recent fields |
zhor recent --since 120 |
alerts the daemon logged over the last two hours |
zhor watch |
the daemon itself, a JSON stream — this is what the widget runs |
Three decisions that matter more than the code:
- A verdict on the first line. The question is always "what is slow", never "give me numbers". The check order is not arbitrary: a memory stall outranks the CPU verdict, because during a stall everything looks busy and the CPU ranking points at the wrong process.
- An exit code.
zhor >/dev/null || collect-diagnosticsneeds no parsing. recent. This is the part a snapshot cannot give you: a freeze gets asked about after it has ended, and by then the machine looks calm. The daemon is running anyway, so it writes alerts to~/.local/state/zhor/events.jsonl(last 2000, self-rotating, deduplicated per process).
There are deliberately no kill/renice subcommands: those already exist and
the report prints pids. Wrapping kill would be extra code and one more way to
hit the wrong thing.
Thresholds are flags; zhor --help lists them all.
zhor watch --interval 3 --cpu-hot 90 --rss-hot 25 --sustain 60 --stall-psi 10Edit them in the command: line in shell.qml. The same file has the pill's
position (pillRight, pillTop), the palette and the colour thresholds at the top.
To catch a permanent hog rather than a build spike, lower the bar and raise the
window: --cpu-hot 50 --sustain 300.
Raw stream without the widget:
zhor watch | head -1 | python3 -m json.tool./test_zhor.pyBare asserts, no framework: CPU delta arithmetic, the stuck-process counter,
alert triggering, shortlist selection, and verdict ordering.
zhor.py measures and decides: /proc + PSI, verdict, alert log, CLI
shell.qml draws and forwards clicks; decides nothing on its own
The split is deliberate: everything that can be computed wrongly lives in the Python, where it is covered by tests and runs without a compositor.
See CHANGELOG.md.
Built end-to-end in a single session by Opus 5 (1M context) — recon, design,
the collector, the QML, the tests, the installer, this README, and the push.
Every number in here was measured on the machine it was written on, not guessed:
the CPU maths was cross-checked against top across 16 processes, and the
uninstall round-trip was diffed byte for byte.
The human contributed the idea, the name, the taste, the refusal to let it autokill anything, one screenshot, and the swearing — including catching that the repo description had gone out in the wrong language.
MIT