Not part of the backend lifecycle (40_backend.sh) -- run by hand during
the live/netinstall boot flow investigation documented in ADR-0009.
kexec-reboot-test.sh/.cfg: tests whether inst.ks= can trigger Anaconda
via kexec from within the running Fedora Cinnamon live session, without
a real firmware reboot. Conclusively fails: the live initrd has no
anaconda dracut module at all.
wlan-netinst-test.cfg: the kickstart used to validate inst.ks=cdrom:/ks.cfg
on a real Fedora netinstall ISO, including the failed WLAN activation
attempts (network --wpakey=, then %pre nmcli) that exposed the missing
iwlwifi driver in that image's initrd.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Replaces the gnome-terminal-launched installer.sh session with a
browser-based kiosk (opt/tuxflotte/kiosk/start-kiosk.sh launching
Epiphany with --profile). Sets German keyboard layout for the
graphical session (vconsole.keymap only covers the text console),
suppresses Epiphany's default-browser prompt, hides desktop icons,
and reduces the panel to a window list + power applet. Also disables
the SELinux troubleshooter desktop notification, and copies the full
live-updates tree (not just etc/) so opt/ payload ships too.
Several Epiphany invocation modes were tried and discarded before
landing on plain --profile=<dir> (see ADR-0004/0006 and platform-docs
history for the rejected --application-mode/--private-instance paths).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The GRUB config searched for the volume by the original Fedora label
(Fedora-Cinn-Live-44), which no longer matches once the ISO is built
with its own Tuxflotte volume ID.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Uses dracut's official 20-apply-live-updates.sh pre-pivot hook: any
/updates/ directory tree present at the top level of the boot media
gets copied into the live root filesystem before pivoting, verbatim,
no initrd or squashfs/EROFS modification needed. build.sh assembles
this tree from scripts/, backends/, config/ (mapped under
/opt/tuxflotte) plus live-updates/etc/ (an XDG autostart entry that
opens a terminal running installer.sh as root, and a sudoers.d drop-in
granting liveuser passwordless sudo).
xorriso preserves the ownership/permissions recorded at map time, so
-chown_r/-chgrp_r 0 on /updates is enough to make sudoers accept the
drop-in as root-owned without needing local root to build the ISO.
This was tried first as a direct EROFS unpack/repack of
/LiveOS/squashfs.img (that file is actually EROFS despite the name on
current Fedora), but a fresh self-built erofs-utils (Debian's
packaged 1.5-1 can't even read this image's on-disk format) hit a
reproducible bug extracting the packed/fragmented inode into a single
corrupt file instead of a directory tree. The dracut hook sidesteps
that path entirely.
Verified end-to-end: booting either Tuxflotte entry reaches the
Cinnamon live desktop, autostarts a terminal, and installer.sh runs
through every module (network, hardware, enrollment, handshake,
device status, Bereitstellungsvorlage selection, commit point, Runtime
Blueprint resolution against the real server, full Fedora backend
lifecycle, storage detection) to a clean exit.
It called log_info/log_warn/error_exit/list_install_disks without
sourcing anything - leftover from before modules were run as
standalone subprocesses via run_module() rather than sourced into
installer.sh. Never noticed because the pipeline never reached this
module in a live run until now. Behavior unchanged: still just lists
disks and warns that partitioning is disabled (Phase 1).
export_connection_profile() crashed the whole installer under set -e/
pipefail when the active NetworkManager connection has no backing
keyfile (common for ephemeral live-boot DHCP connections) - nmcli
returned a non-zero exit status even though the guard below already
handles an empty result gracefully. Every other nmcli call in this
function already has this || true safety net; this one was missing it.
Found via a live end-to-end boot test in Proxmox.
These are run as standalone subprocesses via run_module()/source, which
requires the execute bit. Missing since their initial creation - only
noticed because manually chmod'd copies were used for testing on anode,
never the checked-in files. Would have failed on a fresh checkout.
Blue-on-white (matching pics/tuxflotte-beschriftet-schatten.svg's
brand color rgb(52,101,164)) instead of GRUB's default white-on-black.
The logo image itself doesn't work here: classic gfxterm always paints
an opaque per-character cell background, so a background_image gets
fully covered by color_normal's background and never becomes visible -
confirmed by booting the composited variant in QEMU before dropping it
again. Showing the actual logo needs the GRUB theme/gfxmenu engine,
which this live ISO doesn't ship (empty /boot/grub2/themes, no
x86_64-efi module directory) - not attempted here. The logo fits more
naturally on the Plymouth splash or live-desktop wallpaper instead,
both already confirmed to render correctly.
Replaces the Fedora DVD/netinst base with a live-ISO boot chain
(root=live:CDLABEL=Fedora-Cinn-Live-44 rd.live.image instead of
inst.stage2=/inst.ks=), implementing ADR-0003. The GRUB menu now
offers two entries booting the same live image with a
tuxflotte.mode=auto/interactive kernel cmdline flag instead of a
boot-time kickstart URL - installer.sh will read this flag once it
runs inside the live session.
scripts/build.sh and scripts/extract.sh now extract the source ISO via
`xorriso -osirrox` instead of a loop mount, removing the sudo
dependency for local builds.
Verified end-to-end in QEMU/KVM: GRUB menu renders both entries,
kernel/initrd load correctly, boot proceeds to the live session.
30_runtime_blueprint.sh calls POST /templates/{id}/resolve with the
device_id from the server handshake and the template_id from the
Bereitstellungsvorlage selection, storing the resulting Runtime
Blueprint under /run/tuxflotte/runtime/.
40_backend.sh dispatches to backends/${backend_id}/backend.sh based on
the resolved backend_id and drives the backend_init/validate/
generate_config/launch/postinstall lifecycle from 06-backend-api.md.
backends/fedora/backend.sh implements that lifecycle for Fedora:
backend_generate_config() renders kickstart.tpl via envsubst using the
Runtime Blueprint's installation_directives and the device hostname,
embedding the resolved Merkmal blueprints as JSON for the (not yet
implemented) Provisioning Agent to apply later. Replaces the old
git-clone-based %post bootstrap. backend_launch()/backend_postinstall()
are Phase 1 stubs pending the live-ISO boot integration (see
platform-docs ADR-0003).
20_profile_selection.sh and 25_installation_confirm.sh still spoke the
old flat profile model (.profiles[], .profile.distribution,
.profile.installer.type) that the provisioning server no longer
returns - it now returns .templates[] with workspace/backend objects
(see provisioning-server's Merkmal/Blueprint/Bereitstellungsvorlage
migration). Both modules now read/write
/run/tuxflotte/assignment/template.json against the current
Bereitstellungsvorlage schema, matching 10-interactive-provisioning-flow.md.
