Fix an ISO install that fails with "failed retrieving file ... from disk" for a package in the offline mirror
Checked the installer source in omacom/omarchy-iso on quattro and the archiso runtime source, and read issue 7704 in full with all eleven comments. Nothing here could be exercised on this machine: this is an installed Omarchy 4.0.2-1 system, not a live ISO, so there is no /run/archiso, no offline mirror and no pacstrap to run, and every claim below is from source or from the thread rather than observed. The bind-mount claim is confirmed from source. _mount_offline_package_cache at configs/airootfs/usr/share/omarchy-iso/orchestrator/phases_impl.py:642 runs mount --bind /var/cache/omarchy/mirror/offline <target>/var/cache/pacman/pkg, called at line 250, and configs/pacman-offline.conf carries Server = file:///var/cache/omarchy/mirror/offline/, so during pacstrap the repository and the validated cache are one directory and a --noconfirm delete unlinks out of the mirror. The installer's own omarchy-install-diagnose-media says the same in its header comment. The recovery path is confirmed right for archiso. I pulled mkinitcpio-archiso 73-1 from a pacman mirror and read usr/lib/initcpio/hooks/archiso: line 370 is _mnt_fs "${fs_img}" "/run/archiso/airootfs", _mnt_fs attaches the image with losetup --find --show --read-only and mounts it -r, and line 374 overlays it with upperdir=/run/archiso/cowspace/${cow_directory}/upperdir, where /run/archiso/cowspace is a tmpfs unless a cow_device is passed. Omarchy's configs/grub/grub.cfg cmdline at lines 57 and 63 passes only archisobasedir, archisosearchuuid and display options, so there is no copytoram and no persistent overlay, which makes the record's "reboot brings it back" correct. The checksum command is right too: configs/profiledef.sh sets install_dir="arch", arch="x86_64" and airootfs_image_type="squashfs", and mkarchiso in archiso 90-1 calls _mkchecksum unconditionally, writing sha512sum airootfs.sfs >airootfs.sha512, so the path and the expected airootfs.sfs: OK both hold. One thing was wrong. The cause presented the bind mount as covering the whole install and used it to explain both quoted symptoms, and it does not. _unmount_offline_package_cache runs in the finally that closes the base-install block, before _prepare_target_setup, which bind-mounts the mirror at its own path inside the target rather than onto the cache, leaving the target's /var/cache/pacman/pkg as the btrfs @pkg subvolume written into fstab at line 848. The nvidia-utils failure happened in "Configuring system" under omarchy-apply-system, which is after that unmount, so from source the deletion should not have reached the mirror. I rewrote that paragraph to confirm the mechanism for the pacstrap window, to say it is not established for the later window, and to keep the reporter's observation with the collaborator's reading of the follow-on failed retrieving 'offline.db' from disk as failing reads. Two smaller corrections to the fix: every 4.0.x release note publishes an ISO sha256, not only v4.0.1 (v4.0.0 9224fab3, v4.0.1 69cbb4e1, v4.0.2 2ef8e624, v4.0.3 03d60bc7), and the fix omitted that the failure screen now diagnoses the medium by itself. omarchy-install-diagnose-media landed 2026-08-23 in omarchy-iso PR #120, is installed 0755 per configs/profiledef.sh:42, and omarchy-install-dashboard calls it unconditionally at line 844, so it is on the 4.0.1, 4.0.2 and 4.0.3 ISOs but not 4.0.0. I added it as the first step along with the one case where it stays silent, which follows from its own sed pattern matching only /mnt/var/cache/pacman/pkg while the nvidia-utils reporter's log line was the in-chroot path. Issue 7704 supports every other claim in the record, including the 1247 and 1249 package counts, the microSD-versus-USB switch, and the SDDM empty-username consequence, and it is still open.
Symptom
Installing from the Omarchy 4.0.x ISO fails part way through. During the "Installing Arch + Omarchy" (pacstrap) stage:
failed retrieving file 'gst-plugin-gtk-1.28.6-1-x86_64.pkg.tar.zst' from disk : Could not open file
ERROR: Failed to install packages to new root
or, on the first attempt in a boot, during "Configuring system":
:: File /var/cache/pacman/pkg/nvidia-utils-610.57.04-1-x86_64.pkg.tar.zst is corrupted
Do you want to delete it? [Y/n] error: failed to commit transaction (invalid or corrupted package)
Failed: /usr/share/omarchy/install/hardware/nvidia.sh
Looking in /var/cache/omarchy/mirror/offline/ from the installer shell, the named package really is absent, and /etc/pacman.conf points at Server = file:///var/cache/omarchy/mirror/offline/. Retrying in the same boot fails the same way every time, and a later retry can add failed retrieving 'offline.db' from disk. The network is up and unrelated. Reported on the 4.0.0 (IMAGE_VERSION=2026.08.14), 4.0.1 and 4.0.2 ISOs, naming gst-plugin-gtk (three reporters) and nvidia-utils (one).
Cause
The ISO is not missing the package. A collaborator checked the published 4.0.0 and 4.0.1 images against their own signatures and manifests and found every package in the offline mirror present with a matching sha256 (1247 and 1249 packages respectively). The failure happens between that file and pacman reading it: a bad download, a bad write to the stick, or a bad read off it.
During the pacstrap stage the offline mirror directory is bind-mounted onto pacman's package cache, so the repository pacman fetches from and the cache it validates are the same directory. _mount_offline_package_cache in the ISO installer runs mount --bind /var/cache/omarchy/mirror/offline <target>/var/cache/pacman/pkg, which is done to avoid copying several GiB twice. When one package fails its checksum, pacman's "File ... is corrupted. Do you want to delete it?" is answered yes by --noconfirm and the file is unlinked out of the mirror. The collaborator reproduced that on a clean VM with pacman 7.1.0: intact, the package installs. With 64 bytes altered, pacman reports invalid or corrupted package (checksum) and deletes it, and the next attempt reports the Could not open file error this issue was filed about. The deletion lands in the live session's RAM overlay, not on the stick, because archiso mounts the root image read-only and layers a tmpfs over it, and no Omarchy boot entry asks for a persistent overlay. That is why every retry in the same boot fails on the missing file and a reboot brings it back.
That bind is released before the "Configuring system" stage, so the two symptoms above do not share one mechanism. The installer unmounts the mirror from the target cache in the finally that closes the base-install block, then bind-mounts the mirror at its own path inside the target instead, leaving the target's /var/cache/pacman/pkg as its own btrfs @pkg subvolume. A package pacman deletes during omarchy-apply-system is therefore a copy in that subvolume, and from source it should not remove anything from the mirror. The nvidia-utils reporter on the 4.0.2 ISO nevertheless found the mirror missing the package on the retry, followed by failed retrieving 'offline.db' from disk, and the collaborator's reading of that second error is that the reads themselves had started failing rather than one package having been deleted. Treat a mirror-path error after a "Configuring system" failure as a sign of a failing medium, not as the deletion cascade. What the nvidia-utils report does settle is that nothing here is specific to gst-plugin-gtk.
Two reporters confirmed the media side. One was booting from a microSD card in a USB reader on a USB-C dongle: sha512sum -c airootfs.sha512 reported a mismatch, and the install succeeded after writing the same ISO to a real USB drive. The reporter who opened the issue was on the same microSD-plus-adapter setup.
One reporter also found that when the abort lands in "Configuring system", the target is left half configured: the login step and the post-install step that replaces the offline pacman.conf never run, and the first boot reached an SDDM greeter that rejected the password because the theme submitted an empty username. Treat a failed install as one to redo, not one to boot.
Fix
Read the failure screen first. On the 4.0.1 and later ISOs the installer runs omarchy-install-diagnose-media itself when a step fails and prints a verdict that names the medium, so the answer is often already on screen. It is silent in one case worth knowing: it only matches a corrupt-package line under /mnt/var/cache/pacman/pkg, so a failure logged from inside the chroot during "Configuring system" prints nothing. The 4.0.0 ISO predates the tool.
Then check the medium, from a shell in the live session on a fresh boot, before starting the installer. This re-reads the whole root image off the stick and takes a few minutes:
cd /run/archiso/bootmnt/arch/x86_64 && sha512sum -c airootfs.sha512
A mismatch means the copy on the stick is wrong. Re-download the ISO, check it against the sha256 published in that version's release notes (every 4.0.x release publishes one beside the download link), and write it to a different medium. A plain USB drive worked where a microSD card in a USB reader did not (confirmed by the reporter who switched, and it is the setup the original reporter was on too).
OK means the stick is sound. Reboot between attempts, because a package deleted during pacstrap stays deleted for the rest of that boot, and quote the first error after a fresh boot, since failed retrieving file ... Could not open file is only ever the wake of an earlier failure. /var/log/omarchy-install.log in the live session holds the full text.
To retry without rebooting, put the deleted package back from the read-only image, which archiso keeps mounted at /run/archiso/airootfs for the whole session. The collaborator tested this arrangement on a VM, no reporter has confirmed it on real media:
sudo cp /run/archiso/airootfs/var/cache/omarchy/mirror/offline/gst-plugin-gtk-*.pkg.tar.zst \
/var/cache/omarchy/mirror/offline/
Substitute whichever package the error named, then start the installer again. That cp is also a test, because it re-reads the same bytes off the same stick. It succeeds and the install gets past the package: a one-off misread. The install rejects the same package again with invalid or corrupted package (checksum): the bytes on the stick are wrong, re-download and re-flash. The cp fails with an input/output error: the medium itself is failing, use another port or another stick.
Which package is named after a fresh reboot is a useful signal. The same package again means a stable fault in your copy (bad download or bad write). A different package each time means the bytes are fine and the reads are not (port, cable, card reader, or memory).
Note that the 4.0.0 ISO carries gst-plugin-gtk-1.28.6-1 and 4.0.1 carries -2, so the version in the error tells you which image you booted.
Verify
sha512sum -c airootfs.sha512 prints airootfs.sfs: OK, and the installer runs through "Installing Arch + Omarchy" and "Configuring system" without naming a package. On a stick that failed the check, the same ISO written to a different USB drive installed cleanly.
Sources
- https://github.com/omacom/omarchy/issues/7704
- https://github.com/omacom/omarchy-iso/blob/quattro/configs/airootfs/usr/share/omarchy-iso/orchestrator/phases_impl.py
- https://github.com/omacom/omarchy-iso/blob/quattro/configs/pacman-offline.conf
- https://github.com/omacom/omarchy-iso/blob/quattro/configs/profiledef.sh
- https://github.com/omacom/omarchy-iso/blob/quattro/configs/grub/grub.cfg
- https://github.com/omacom/omarchy-iso/blob/quattro/configs/airootfs/usr/local/bin/omarchy-install-diagnose-media
- https://github.com/omacom/omarchy-iso/blob/quattro/configs/airootfs/usr/local/bin/omarchy-install-dashboard
- https://archlinux.org/packages/extra/any/mkinitcpio-archiso/
- https://archlinux.org/packages/extra/any/archiso/
- https://github.com/omacom/omarchy/releases/tag/v4.0.0
- https://github.com/omacom/omarchy/releases/tag/v4.0.1
- https://github.com/omacom/omarchy/releases/tag/v4.0.2
- https://github.com/omacom/omarchy/releases/tag/v4.0.3