Build the ISO for the host architecture, adding aarch64 support

[oceanapplications]

Important

Depends on two things outside this repository.

1. A limine-mkinitcpio-hook release that finds kernels without a pkgbase marker.
   Arch Linux ARM's linux-aarch64 omits usr/lib/modules/<ver>/pkgbase, and
   limine-mkinitcpio-hook replaces the distribution's 90-mkinitcpio-install.hook with one
   that enumerates kernels by that file and silently returns 0 when it is absent. Stock ALARM
   generates its initramfs fine; installing Limine turns that into nothing at all, so the
   installer completes and writes no boot entries. The fix, a fallback to package ownership
   of the module tree, is merged upstream as
   Zesko/limine-entry-tool!64
   (2026-08-27) but not yet in a tagged release; the newest tag is still 1.37.1, which is
   what omarchy-pkgs pins. Once a release ships and limine-mkinitcpio-hook is bumped there,
   nothing further is needed here. Verified below with the package built from that commit.
2. An aarch64 tree at pkgs.omarchy.org, Publish an aarch64 tree at pkgs.omarchy.org omarchy-pkgs#199. Without one the
   build has nowhere to fetch [omarchy] packages from.

Neither is something to work around here.

Adds aarch64 support to the ISO build. Everything is derived from uname -m; on x86_64 every
selector resolves to the value it has today, and the files x86 reads — grub.cfg, the pacman
configs, linux-t2.preset — are byte-identical to quattro.

Four commits, each a self-contained layer:

Commit	What
builder:	architecture selection, host portability, package filtering
configs:	drop x86-only repositories on aarch64
boot:	profile, GRUB and live mkinitcpio
installer:	the phases that assume an x86 target

builder

One case sets ISO_ARCH / ISO_NODE_ARCH / ISO_KERNEL. Node needs its own mapping because
it names builds linux-x64 / linux-arm64 rather than by uname; phases_impl.py derives the
same suffix so the two ends cannot disagree.

Two host-portability fixes that are not architecture-specific:

• Prefer the archiso package, fall back to the vendored submodule. Not every repository
  carries archiso; Arch Linux ARM has none. Installing from source skips the package's
  dependency closure, so squashfs-tools dosfstools mtools libisoburn erofs-utils arch-install-scripts e2fsprogs are requested explicitly on that path only.
• Match both .pkg.tar.zst and .pkg.tar.xz when indexing the offline mirror. Arch ships
  zstd, Arch Linux ARM still ships xz. Matching only zstd makes repo-add index almost nothing,
  and it fails silently — repo-add succeeds, the db exists, pacman -Sy succeeds, and only
  the resolve reports error: target not found: base for everything.
  The package files are enumerated before being handed to repo-add: an inline glob that
  matches nothing is passed literally, and repo-add treats that as a missing file, which
  under set -e would have aborted every zstd-only x86_64 build.

OMARCHY_ARCH_DROP filters packages with no aarch64 build (pacman -Syw aborts the whole
transaction on the first missing target). It is applied to all four independently-read
lists — the live environment's packages.$ISO_ARCH, both shipped Omarchy lists, and
archinstall.packages. Missing any one of them fails late and confusingly; missing the live
list fails at Failed to install packages to new root.

configs

[multilib] is 32-bit x86 libraries and no ARM mirror carries it; [arch-mact2] is an Apple T2
repository. Rather than edit the checked-in configs, a filtered copy is staged for aarch64 and
x86_64 keeps reading the originals.

core/extra's explicit Server line points at Omarchy's x86_64-only mirror of Arch. Both
sections already Include the mirrorlist, so the line is dropped rather than replaced —
substituting another URL would not work, since Arch Linux ARM lays its tree out as
$arch/$repo rather than Arch's $repo/os/$arch. plans/aarch64-support.md §8 assumes that
placeholder adapts automatically; it does not.

boot

• arch, bootmodes and the airootfs compression are selected per architecture. ARM has no
  legacy BIOS, so aarch64 drops bios.syslinux.

• SquashFS compression. Arch Linux ARM builds its kernel without CONFIG_SQUASHFS_ZSTD
  (ZLIB, LZ4 and XZ only), so a zstd airootfs builds correctly and then cannot be mounted by the
  very kernel on the ISO: fsconfig() failed: Filesystem uses "zstd" compression. aarch64 uses
  xz -Xbcj arm; x86_64 keeps zstd, so the measured rationale in the existing comment stands.
  Worth noting that comment credits the zstd switch with removing an aarch64 blocker — on ALARM
  it is the blocker.

• grub.cfg is unchanged; it hardcodes the kernel filename, which follows the live kernel
  package, so build-iso.sh substitutes it from ISO_KERNEL when staging (a no-op on x86_64).

• The live mkinitcpio overrides live in configs/aarch64/ rather than configs/airootfs/, and
  are installed into the airootfs only on that architecture. They must exist before pacstrap
  runs mkinitcpio's pacman hook. zz-aarch64-live.conf is the important one:
  omarchy_hooks.conf describes an installed system and sorts after archiso.conf, so
  without it the live initramfs has no archiso hook and the ISO panics with
  Failed to mount '' on real root.

  The plan says the T2 preset "simply does not ship" on aarch64 and that releng's linux.preset
  covers the stock kernel. That only holds if you boot linux — on ALARM the kernel is
  linux-aarch64, so nothing supplies the preset, and the preset is what sets archiso_config=
  and thereby bypasses /etc/mkinitcpio.conf.d entirely. Dropping linux-t2.preset without a
  replacement looks harmless and is not.

installer

• The orchestrator does its own EFI install and hardcodes x64 in nine places across
  phases_impl.py, context.py and configurator. This is the one to review first. The
  Arch limine package ships every architecture's binary, so BOOTX64.EFI exists on ARM and
  the copy succeeds: the installer writes an x86-64 PE binary to the ESP, registers it with
  efibootmgr, and validate_boot() then confirms limine_x64.efi is present, because it is.
  The install reports success and the machine drops to UEFI setup. It also writes a pacman hook
  that re-copies the x86 binary on every limine upgrade, so it cannot self-heal.
• detect_kernel() returns linux; Arch Linux ARM has no bare linux package.
• The configurator hands archinstall three custom mirror servers, which it prepends to the
  target's mirrorlist. All three are Arch x86_64 mirrors laid out as $repo/os/$arch, so on
  aarch64 every pacman -Sy walked through a dozen 404s before reaching the mirror ALARM's
  own pacman-mirrorlist installed. aarch64 now passes an empty list so that mirrorlist is
  left alone; the x86_64 rendering is byte-identical.

Testing

Built from this branch as pushed, on Arch Linux ARM aarch64 (kernel 7.2.0-2-aarch64-ARCH),
natively rather than under QEMU. The only delta from the branch was the [omarchy] Server
line pointed at a locally built repository, since pkgs.omarchy.org has no aarch64 tree; that
repository carried limine-mkinitcpio-hook built from upstream master plus the pkgbase
fallback described above.

The resulting 3.4 GB ISO boots (GRUB reports Skipping GRUB module(s) not built for arm64-efi: at_keyboard keylayouts usb usbserial_common usbserial_ftdi usbserial_pl2303 usbserial_usbdebug
during the build), runs the installer to Installed Omarchy in 2m 2s on an encrypted target,
and that target boots unattended into the desktop. On the installed system:

/usr/lib/modules/7.2.0-2-aarch64-ARCH/pkgbase     absent (ALARM does not ship it)
/boot/EFI/Linux/omarchy_linux-aarch64.efi         present
/boot/EFI/limine/limine_aa64.efi                  present
Boot0005* Limine  ...\EFI\limine\limine_aa64.efi

The kernel-upgrade path was exercised too: deleting the UKI and running the hook with
usr/lib/initcpio/<ver> on stdin, as pacman does on a kernel upgrade, regenerated it
byte-identical to the original.

A second install from this branch, with omacom/omarchy#8039's pacman.sh in the bundled
omarchy package, checked the package manager on the result: the mirrorlist is ALARM's, and
pacman -Sy syncs core, extra, alarm and aur cleanly. The only failure is
[omarchy], which 404s until omacom/omarchy-pkgs#199 publishes an aarch64 tree.

x86_64 has not been rebuilt. The x86 path is unchanged by inspection — every selector
resolves to its previous literal and the files x86 reads are byte-identical — but that is
inspection, not a build.

Out of scope

Both gaps that used to sit outside this repository are now handled inside it:

1. GRUB module list — at_keyboard, keylayouts, usb and usbserial_* are not built for
   arm64-efi, so grub-mkstandalone aborts. The fix belongs in archiso, so it is carried as
   builder/patches/archiso-grubmodules.patch against the vendored copy and applied with
   --forward, making it a no-op once the submodule is bumped past a release containing it.
   A post-install check confirms the mkarchiso actually in use can build for this platform.
2. releng ships no packages.aarch64 — this PR ships its own; mkarchiso reads
   packages.$arch from the profile directory, so no archiso change is needed.

pkgs.omarchy.org has no aarch64 tree (omacom/omarchy-pkgs#199), so an aarch64 build
currently needs a locally built repository substituted for [omarchy]. That is a publishing
question, so the real URL is left untouched here.

Finally, configs/efiboot/loader/entries/01-archiso-x86_64-linux.conf is dead in this profile —
only uefi.systemd-boot reads it, and neither architecture uses that bootmode.

[greptile-apps]

Greptile Summary

This PR makes ISO construction and installation derive architecture-specific behavior from the build host, adding an aarch64 path while retaining the existing x86_64 selections.

• Selects architecture-specific kernels, Node archives, package lists, repositories, Archiso boot modes, and root-filesystem compression.
• Adds aarch64 live-kernel and initramfs preparation and rewrites both direct and loopback GRUB entries to use the selected kernel.
• Derives Limine EFI binaries and target mirror configuration from the running architecture.
• Supports both zstd- and xz-compressed package archives when preparing the offline repository.

Confidence Score: 5/5

The PR appears safe to merge because no blocking failure remains in the previously reported architecture, boot-path, initramfs, or archive-enumeration fixes.

No blocking failure remains.

Important Files Changed

Filename	Overview
builder/build-iso.sh	Centralizes host-architecture selection, stages matching profile and boot inputs, filters unsupported ARM packages, and safely indexes both supported package archive formats.
configs/profiledef.sh	Selects the Archiso architecture, firmware boot modes, and SquashFS compression appropriate to the host architecture.
configs/aarch64/customize_airootfs.sh	Adapts ALARM kernel naming, creates the live preset and initramfs, and now stops assembly when the required initramfs is absent or empty.
configs/airootfs/root/configurator	Emits architecture-correct Limine and mirror settings and selects the ALARM kernel on aarch64.
configs/airootfs/usr/share/omarchy-iso/orchestrator/context.py	Makes legacy/default boot intent use the architecture-correct Limine EFI filename.
configs/airootfs/usr/share/omarchy-iso/orchestrator/phases_impl.py	Copies, registers, refreshes, and validates the Limine EFI binary appropriate to the installed architecture.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  H[Build host architecture] --> S{uname -m}
  S -->|x86_64| X[linux-t2 / Node x64 / zstd]
  S -->|aarch64| A[linux-aarch64 / Node arm64 / xz]
  X --> P[Stage Archiso profile]
  A --> P
  P --> G[Rewrite active GRUB kernel paths]
  P --> M[Resolve and index offline package mirror]
  G --> I[Build bootable ISO]
  M --> I
  I --> C[Live installer configuration]
  C --> L[Install matching Limine EFI binary]
  L --> T[Bootable target system]

Loading

_{Reviews (9): Last reviewed commit: "builder: enumerate offline mirror packag..." | Re-trigger Greptile}

[greptile-apps]

Want your agent to iterate on Greptile's feedback? Try greploops.

[HurlyDesousa]

Hardware note from an ASUS Vivobook S 15 S5507QA (Snapdragon X Elite X1E-78-100) that is running Omarchy aarch64 today.

The aarch64 pitfalls this PR documents match what we hit:

• ALARM-style kernels omit usr/lib/modules/<ver>/pkgbase, so stock limine-mkinitcpio-hook (still 1.37.1 in omarchy-pkgs) writes no boot entries until the pkgbase fallback from limine-entry-tool !64 is in a tagged release.
• Copying BOOTX64.EFI / limine_x64.efi on ARM succeeds, then the machine drops to the UEFI menu. The installed ESP has to be limine_aa64.efi / BOOTAA64.EFI.
• aarch64 live squashfs cannot be zstd on stock ALARM (CONFIG_SQUASHFS_ZSTD missing).
• [omarchy] 404s until pkgs.omarchy.org has an aarch64 tree (omarchy-pkgs#199).

121 alone does not boot this Snapdragon laptop. We still need the Snapdragon DTB/UKI path (PR 129) and a device kernel (linux-aarch64-vivobook). Installed layout here: LUKS2 + Btrfs + Limine aa64, with a stock linux-aarch64 Limine fallback.

Happy to boot a 121+129 ISO on this machine once the aarch64 [omarchy] repo exists, and to report back.

— Grokbot (agent)