## What happens

The Omarchy shell (`quickshell -n -p /usr/share/omarchy/shell`) accumulates wl_shm
buffer file descriptors and RSS for as long as it runs, with an idle desktop and no
user interaction. It is also the dominant driver of a matching fd leak in Hyprland,
which ends the session in a compositor abort.

My Hyprland session died twice this way on 2026-08-24 (20:51 and 22:35 PDT).

## Expected

A shell sitting idle should hold a steady number of buffers.

## Evidence — shell side

Single quickshell process, no restarts, 54 minutes old:

```
quickshell  pid=56066  uptime=3251s  fds=5122  RSS=623 MB
```

4,900 of those 5,122 fds are read-only handles to deleted `/dev/shm/<uuid>` segments.
Two segments dominate and keep growing:

```
inode 23 : 2005 fds held by quickshell
inode  7 :  498 fds held by quickshell
```

The shell's own journal is clean — 53 lines this boot, config loaded twice, no errors
that account for this. It leaks silently while behaving normally.

Worth noting: nothing in the environment forces software rendering
(`QT_QPA_PLATFORM=wayland;xcb`), so I do not have an explanation for why the shell is
on the wl_shm path rather than dmabuf on this machine. That may be the actual root —
if the shell used the GPU path there would be no shm churn to leak.

I could not separate quickshell-upstream behaviour from Omarchy's QML without
instrumenting the shell, so I am not asserting which side owns it.

## Evidence — compositor side, and why the session dies

Hyprland accumulates ~8x more fds for the same segments and never releases them,
including after the client exits:

```
uptime         : 4425s
total fds      : 44162 / 524288
wl_shm fds     : 43024
  client exited: 17549      (across 213 dead segments)

inode 23:  Hyprland = 16571 fds   vs   quickshell = 2005 fds
```

Idle growth, nothing animating: **+1091 wl_shm fds in 30 seconds.**

That Hyprland half is a separate upstream bug. It was already reported as
hyprwm/Hyprland#14662 (2026-05-18) and auto-closed nine seconds later by the
no-user-issues bot, never triaged; I am raising it again in their Discussions with the
compositor-side evidence. Note #14662 identifies the trigger precisely as
**transparent zwlr_layer_shell_v1 surfaces using wl_shm** -- which is what the Omarchy
bar is. It is here because the shell is what drives it hardest, and because the
combination is what kills an Omarchy session.

## The failure chain

1. Shell allocates wl_shm buffers continuously; neither side releases them
2. Hyprland walks its fd table to EMFILE — confirmed in both crash reports:
   `GPU:  <pipe(pipefd) failed with24>` (errno 24 = EMFILE)
3. It can no longer dup dmabuf/sync fds:
   `eglCreateImageKHR ... could not bind to DMA buffer`
4. Clients stall; the ANR watchdog fires and spawns `hyprland-dialog`
5. Those dialogs cannot be handed a DRM fd either — they get `-1` and abort:

```
m_drmState = {fd = -1, nodeName = ""}
#3 Hyprtoolkit::COpenGLRenderer::COpenGLRenderer  hyprtoolkit-0.5.4/src/renderer/gl/OpenGL.cpp:421
     RASSERT(false, "Couldn't open a gbm fd");
#5 Hyprtoolkit::IBackend::create                  hyprtoolkit-0.5.4/src/core/Backend.cpp:109
#6 main                                           hyprland-guiutils-0.2.2/utils/dialog/src/main.cpp:108
```

6. The next input-capture session cannot allocate its EIS backend fd; libeis asserts
   and takes the compositor with it:

```
assertion "eis->backend" failed, ../libei/src/libeis-fd.c:78, eis_backend_fd_add_client
  <- CInputCaptureResource::CInputCaptureResource  src/protocols/InputCapture.cpp:46
```

uwsm relaunches Hyprland immediately, so from the user's side the whole desktop
flashes and comes back with everything closed.

## Reproduce

Log in and leave the desktop idle. Watch both counters climb:

```bash
HP=$(pgrep -x Hyprland | head -1); QS=$(pgrep -x quickshell | head -1)
watch -n30 "echo Hyprland=\$(ls /proc/$HP/fd|wc -l) quickshell=\$(ls /proc/$QS/fd|wc -l)"
```

## Mitigation available to users today

- `omarchy-restart-shell` reclaims the shell's fds but **not** the compositor's —
  those are only released when Hyprland restarts
- The only full reset is logging out and back in
- No configuration change avoids it

## System

```
Omarchy        4.0.1-1
Hyprland       0.56.2-1        quickshell     0.3.1-1
hyprtoolkit    0.5.4-4         hyprland-guiutils 0.2.2-2
aquamarine     0.14.0-2        libei          1.6.0-1
Kernel         7.1.8-arch1-3
GPU            NVIDIA GeForce RTX 3080 (GA102), nvidia-open-dkms 610.57.04-1
mesa           1:26.2.1-1  (1:26.1.8-1 at the time of both crashes)
Monitors       2x 2048x1152 @ scale 2 (DP-1, DP-2)
Machine        ASUS ROG STRIX B550-I, Ryzen 7 5800X, 31 GiB RAM
```

Upstream references:
- hyprwm/Hyprland#14662 -- same defect, client-side symptom, auto-closed untriaged
- hyprwm/hyprtoolkit -- the `hyprland-dialog` abort, filed separately

Full diagnostics available from `omarchy debug --no-sudo --print` on request.

---
Filed by Claude Opus 5 via Claude Code.