- koha-shell -c wraps the command in single quotes, which broke the env
assignments; use runuser + env instead and verify tables were created
- koha-create requires mpm_itk, so keep it and grant the caps it needs
- stop any Apache started by koha-create before running it in foreground
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013wi3ZSot4XbUViwRFgLpRf
- Run Apache as the instance user instead of mpm_itk AssignUserID
- Add Listen for the staff port; start zebra/indexer/worker/plack explicitly
- Tolerate koha-plack --enable returning non-zero under set -e
- install.sh: only fail on a crash loop, not a single restart
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013wi3ZSot4XbUViwRFgLpRf
Compose stack (MariaDB + Koha built from the Koha project's Debian
packages), headless bootstrap replacing the web installer, seeded item
types and LoC Z39.50 server, and an install.sh that installs Docker,
generates passwords and waits for the site. End-to-end testing in progress.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013wi3ZSot4XbUViwRFgLpRf
Unattended installs on an --offline ISO cap the online attempt (default 40 min)
and finish from the bundle on timeout or failure. Interactive retry menu gains
[O] to install from the bundle. NodeSource curl gets timeouts.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A97KUggSE3rBL8veDM5Mr4
kiosk-wifi-helper gains region-status/region-set. The country is persisted as
a cfg80211 regdom module option and in the WiFi netplan file; locale and
keyboard go through localectl, validated against what's installed.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A97KUggSE3rBL8veDM5Mr4
Without a country Linux uses the restrictive world domain: networks show up
in scans but the adapter can't join them. The menu asks (default from the
existing config or system locale); the web helper uses the same default.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A97KUggSE3rBL8veDM5Mr4
- Menu asks which adapter when several exist (shows USB/PCI); the web UI
helper prefers the adapter already configured.
- Stale-entry cleanup also marks remaining installer wifis optional so a
dead WiFi entry can't hold up boot.
- After a failed connect, read wpa_supplicant's log to say wrong password /
network not found / rejected.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A97KUggSE3rBL8veDM5Mr4
netplan apply waits ~90s for every wifis: interface in every file and then
fails if one doesn't exist (e.g. installer config naming wlan0 while the
adapter is wlp1s0). Remove such entries (backup kept as *.kiosk-bak) before
applying, in both the menu and kiosk-wifi-helper.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A97KUggSE3rBL8veDM5Mr4
The WiFi screen now opens with Ctrl+Alt+Super+W, which is easier to
press than the five-key chord. The Shift+W+S+F+H chord and
Ctrl+Alt+Shift+Super+W are removed. Fn isn't part of the combo: most
keyboards never send it to the computer.
With password protection on, the screen asks for the unlock code: the
PIN pad for a PIN, a password field for a password. It shares the
lock screen's wrong-code slowdown. With protection off, it falls back
to the hidden-sites PIN as before. The lock screen's check now goes
through the same checkCode() helper.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
.jitsi-pin now holds "sha256:<hex>" (or NOPIN) instead of the digits.
The menu writes the hashed form. A plain-digits file still works and
is converted by the kiosk app at startup.
The PIN is now checked in the main process instead of the PIN pad's
renderer, using the lock screen's slowdown, now factored out
(createCodeThrottle): 5 wrong in a row refuse entry for 30s, doubling
up to 5 minutes. The F10 PIN pad and the WiFi screen share one count.
The WiFi screen's close-after-5-tries is replaced by this. The PIN pad
no longer shows the default PIN on screen, and the menu no longer
echoes the PIN as it's typed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
The unlock code and the boot lock code can each be a password or a
4-8 digit PIN. A PIN gets an on-screen number pad, so the lock works on
a touch-only kiosk. The boot lock can use the unlock code or its own.
These are the kiosk's own codes, set in Core Settings -> Password
Protection & Lockout or on the Web UI's Lockout page, never the admin
(sudo) password.
The lock screen moves from an inline data: URL to lockscreen.html. The
code is now hashed and compared in the main process, which also slows
down guessing: after every 5 wrong codes in a row, input is refused for
30s, doubling up to 5 minutes. Existing configs behave as before:
missing fields mean a password, and the same code at boot.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
With the virtual consoles turned off there was no way to join a WiFi
network at the kiosk itself. Holding Shift+W+S+F+H together (or
Ctrl+Alt+Shift+Super+W) now opens a PIN-protected WiFi screen: scan,
pick or type a network, enter the password, connect.
The root side is a small helper (kiosk-wifi-helper: status, scan,
connect only) that the kiosk user may run through a validated
sudoers.d rule. It writes the same netplan file as Core Settings ->
WiFi and puts the previous one back if the new network gives no
address. Installed by provisioning and Advanced -> Upgrade, removed by
Complete Uninstall.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
Easy2Boot's UEFI menu returned straight to the menu for the new kiosk ISO
(a 3.5 GB copy on a well-used drive). Some multiboot modes need the ISO in
one unbroken piece, and Linux has no defragmenter for NTFS/FAT32/exFAT -
Easy2Boot's own Linux helpers are a 32-bit udefrag needing i386 libraries
and a FAT32-only perl script.
e2b-contig.sh rewrites just the one file, using only ordinary file
operations - never filesystem structures - so the worst case is "still
not contiguous", never a damaged drive:
1. fill the drive's free space with 4 MB filler files (plugs every gap;
a plain copy, or fallocate where it exists at all, just fills the
gaps again - measured on ntfs-3g: 7-8 pieces either way);
2. map the fillers (filefrag -e, which also works on ntfs-3g via FIBMAP)
and find the longest physically unbroken run of single-piece ones;
3. free just enough of that run and write the copy - the only free space
left - then delete the fillers;
4. verify the copy's SHA-256 against the original and rename it into
place (same directory, so the data doesn't move).
If there's no run big enough while the original is on the drive,
--move-out parks it on this computer (checksummed), retries, and puts it
back verified if that still fails. --check only reports. Interrupts and
failures remove fillers and temp copies.
Tested on NTFS images through ntfs-3g: a 60 MB file in 7 pieces became 1
piece with identical content, no leftovers, free space unchanged (27 s for
~900 MB free); drives with no big enough free run (with and without
--move-out) report so and leave the original byte-identical, parked copy
cleaned up. FAT32/exFAT through their FUSE drivers can't report layout,
which the tool detects and stops on, naming the kernel driver to remount
with (vfat/exfat/ntfs3); the kernel drivers aren't available in the test
environment, so those paths are untested.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
Reported: Core Settings -> WiFi "just returns to the core settings menu".
run_menu clears the screen the moment a handler returns, and every early
exit in wifi_menu returned without a pause - so whatever went wrong
(no adapter found, interface wouldn't come up, ...) was wiped before it
could be read.
- wifi_menu now runs the wizard (renamed wifi_configure) and always
pauses afterwards, success or failure.
- Adapter detection by type (/sys/class/net/*/wireless or phy80211),
falling back to wl* - not every adapter is named wl*. When none is
found, lists the interfaces that do exist.
- Clears a software rfkill block via sysfs before `ip link set up`
(rfkill isn't installed on a server); warns on a hard block; shows
ip's actual error if bringing the link up still fails.
- Writes its own $NETPLAN_DIR/60-kiosk-wifi.yaml instead of overwriting
the first *.yaml found - on a fresh install that's the installer's
50-cloud-init.yaml with the wired config, which it silently dropped.
- SSID/password written as YAML single-quoted scalars: a password with
" or \ produced invalid YAML before.
- Success = the WiFi interface itself has an IPv4 address (waits up to
30s), not "the machine has some IP". The SSH revert watchdog uses the
same test instead of pinging 8.8.8.8, which on a network without
internet reverted every working connection; with no previous WiFi file
it reverts by removing ours.
Tested with ip/netplan/sudo stubbed and netplan files in a scratch dir:
installer's wired file untouched, own file written, SSID
`Bob's "Café" #2` and password `p@ss"wo'rd\\:#x` round-trip through a
YAML parser exactly, no-address case reports instead of claiming
success, early errors stay on screen.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
Reported on a 26.04 offline install, first-boot step 4:
chown: cannot dereference '/home/kiosk/kiosk-app/*.js': No such file or directory
(same for *.mjs, *.html). `sudo chown ... "$KIOSK_DIR"/*.js` has the glob
expanded by the calling (admin) shell before sudo runs; the kiosk user's
home is created mode 750, so the admin can't list it, the pattern stays
literal, and chown fails - stopping the install under set -e.
- provision_install_app: chown the copied top-level files with
`sudo find "$KIOSK_DIR" -maxdepth 1 -type f -exec chown ...` (root does
the listing). Top level only, so chrome-sandbox inside node_modules
keeps its root ownership.
- Swept lib/ and menus/ for the same class of bug: Asterisk Intercom's
baresip_is_installed and status read $KIOSK_HOME/.baresip/accounts
without sudo, so they always reported "not configured". Now sudo test
/ sudo head. Everything else that touches the kiosk home (config.json,
kiosk-app checks) already goes through sudo -u kiosk.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
Reported on a 26.04 offline install, first-boot step 2:
useradd: group 'netdev' does not exist
useradd refuses to create the user at all when any -G group is missing,
which stopped provisioning. Ubuntu 26.04 Server's base system has audio,
video, input and plugdev but no netdev (checked against the 26.04.1
ISO's /etc/group).
provision_create_kiosk_user now `groupadd -f -r`s each supplementary
group first (a no-op when it exists), then passes the same list to
useradd.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
A default Ubuntu Server 26.04 install only gets wpasupplicant/iw when
WiFi is configured in the installer's network screen (they're not in the
server manifest otherwise). Without them netplan can't join a WiFi
network and Core Settings -> WiFi stops at "No WiFi tools found" - which
is exactly the offline-first case: install with no network, then use the
kiosk's WiFi menu to connect.
- Add wpasupplicant and iw to PROVISION_APT_PACKAGES, so first boot
installs them in step 1 (from the offline bundle when there's no
internet) - before Core Settings -> WiFi is offered. +1.6 MB in the
bundle.
- wifi_menu installs them itself (via run_with_offline_fallback) if
they're missing, instead of erroring with an apt command to run.
Bundle rebuilt: contains wpasupplicant and iw.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
Reported: installed from the ISO, files present, but the kiosk setup
never appeared on first boot.
Cause: kiosk-firstboot.service had After=cloud-final.service, and on 26.04
cloud-final.service is After=multi-user.target - while multi-user.target,
which wants kiosk-firstboot, is implicitly After= it. An ordering cycle;
systemd breaks it at boot by deleting a job, and kiosk-firstboot's was
the one to go. systemd-analyze verify against the 26.04.1 ISO's own units
reproduces it:
Found ordering cycle: kiosk-firstboot.service/start after
cloud-final.service/start after multi-user.target/start - after
kiosk-firstboot.service
- Drop After=cloud-final.service. The script's `cloud-init status --wait`
goes too: cloud-final waits for multi-user.target, which waits for this
service - it would deadlock. DPkg::Lock::Timeout already covers what it
was guarding against (dpkg lock contention).
- Drop Conflicts=getty@tty1.service: getty.target pulls getty@tty1 into
the same boot transaction, and a conflict between two wanted jobs is
also resolved by dropping one. Before=getty@tty1 alone keeps tty1's
login prompt waiting until this oneshot finishes.
Verified: systemd-analyze verify of kiosk-firstboot + multi-user.target +
cloud-final + cloud-init.target + getty@tty1 + graphical.target in the
26.04.1 server filesystem - old unit reports the cycle, fixed unit is
clean.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
Offline bundle now carries every addon that works on a network without
internet: LMS (Lyrion's own .deb, from the addon's LMS_DEB_URL) +
Squeezelite, Asterisk Intercom (baresip), VNC, WireGuard and the
Emergency Hotspot packages - each list in a variable next to the code
that installs it (SQUEEZELITE_APT_PACKAGES, ASTERISK_INTERCOM_APT_PACKAGES,
VNC_APT_PACKAGES, WIREGUARD_APT_PACKAGES, EMERGENCY_HOTSPOT_APT_PACKAGES),
read by build-bundle.sh. Each install goes through
run_with_offline_fallback like CUPS, so they install offline from the
normal menus / Web UI. LMS installs as a bundle package when offline
(lms_offline_install_wanted); its online repo/direct-download paths are
unchanged. Tailscale/Netbird deliberately not bundled (vendor scripts,
vendor servers). Bundle: 524 MB (+87 MB).
Emergency hotspot: the boot check pinged only 8.8.8.8, so on a LAN with
no internet the hotspot started on every boot. Now chosen at setup:
"network" (default - router or internet must answer; a router that
ignores ping still counts if it answered ARP), "internet" (old check),
or "host" (a specific address). Three tries 10s apart. Existing scripts
keep the old check until Reconfigure.
New Core Settings -> Time Server (NTP) (menus/time_sync.sh): point
systemd-timesyncd at LAN time server(s) via a timesyncd.conf.d drop-in,
back to Ubuntu's defaults, or set the clock by hand (and the RTC) for a
network with no time server. Complete Uninstall resets it.
Also: lms_service_name used `systemctl list-unit-files | grep -q`, the
same pipefail/SIGPIPE pattern as the earlier cups_is_installed bug; now
`systemctl cat`.
Tested in a network-less chroot of the 26.04.1 server filesystem: kiosk
base, then action_wireguard_install and action_install_lms (real
functions) and the Squeezelite/Intercom/VNC/hotspot package sets all
installed from the bundle, apt left on normal sources after each.
Hotspot script generated for all three modes (plus an invalid address,
which falls back to "network") and its check run against simulated
networks: LAN-only with router answering ping or only ARP -> no hotspot
in "network" mode; nothing reachable -> hotspot. Time Server actions
driven with timedatectl stubbed: drop-in contents, input validation,
manual set-time sequence, reset removes the drop-in.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
PDFs in a slideshow folder now play alongside images and Word documents:
a one-page PDF is shown whole, a longer one fits the width and scrolls
through its pages during docInterval (first 30 pages - rendering a huge
PDF for a few seconds on screen isn't worth the time and memory).
- kiosk-app/pdf-render.mjs renders pages to canvases with pdf.js
(pdfjs-dist ^6.3.289, new kiosk-app dependency): no viewer toolbar, and
the slideshow's existing auto-scroll works on it. pdf.js >= 4.2.67 is
required (CVE-2024-4367: older versions could run script embedded in a
crafted PDF); isEvalSupported:false too. It's a module because pdf.js
only ships as one - confirmed to load (with its worker) from file:// in
Electron 44.
- Hands pdf.js a copy of the bytes: it transfers the buffer it's given,
which would break the cached copy the next time round.
- slideshow.js lists .pdf as a document type; the title-request handler
serves any listed non-image item (still only from that folder).
- provision_install_app now copies kiosk-app/*.mjs as well - without it
pdf-render.mjs would never reach an installed kiosk.
Tested with the real main.js on Electron 44.5.1 under Xvfb + Openbox,
app files copied with provision_install_app's own glob: a photo, a
one-page PDF, a three-page PDF and a .docx played in order, twice round;
the one-page PDF shown whole, the three-page one scrolled through all
pages. Offline bundle rebuilt with pdfjs-dist (437 MB total, +22 MB).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
A slideshow folder now plays .docx files alongside images: a one-page
document is shown whole (letterboxed like a photo); a longer one fills
the width and scrolls smoothly top to bottom during its docInterval
(new slideshow.json key, default 2x interval). Word's ~$ lock files are
skipped; changed documents (by mtime) are re-read on the minute refresh.
- kiosk-app/docx-render.js: the docx-preview rendering now shared by
docview.html and slideshow.html. Each render gets its own class prefix
and keeps its styles in its own container, so two documents crossfading
can't restyle each other. "One page" is judged by content fitting the
paper height, not section count - docx-preview only starts a new
section at explicit page breaks, so a long unbroken document is one
tall section.
- Document bytes are fetched on demand: the page sets its title to a
request, slideshow.js answers via executeJavaScript - only for a .docx
in that slideshow's own current listing (no file-read API exposed to
pages; the preload is shared by every site).
Tested with the real main.js on Electron 44.5.1 under Xvfb + Openbox on a
folder of 2 images + a 1-page and a long .docx (+ a ~$ lock file): order,
crossfades, whole-page fit, scroll to the end, lock file skipped; the
single-.docx site still renders. The request handler answered a listed
.docx and refused /etc/passwd, a listed image, and a .docx outside the
folder.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
A site whose path is a local .docx renders as pages in the kiosk instead
of Chromium offering a download. kiosk-app/docview.js maps the file URL
to docview.html (alongside slideshow.js, through main.js's new
resolveLocalSite) and pushes the file's bytes into the page from the main
process; the page renders them with docx-preview + jszip, both pure-JS
npm dependencies of kiosk-app (so online installs, Advanced -> Upgrade
and the offline bundle all pick them up - no LibreOffice).
- Scaled to the screen width (capped at 160%), scrolls normally.
- Re-checked every minute; a changed file re-renders from the top.
- Word's Symbol/Wingdings private-use bullet glyphs (U+F0B7 etc.), which
render blank without Microsoft fonts, are mapped to Unicode (•, ▪, ...).
- Missing/unreadable/oversized (>50 MB) files and a missing library show
a message instead of a blank page. .doc/.odt aren't supported.
Tested with the real main.js on Electron 44.5.1 under Xvfb + Openbox:
heading, coloured bold run, image, bordered table, bullets and a page
break rendered; wheel scrolling works; replacing the file while running
re-rendered the new document within the refresh. Offline bundle rebuilt
and contains docx-preview and jszip.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
Slideshow: a site whose path is a local folder (/home/kiosk/photos/, or
any existing directory) plays its images full-screen instead of showing
Chromium's directory listing. kiosk-app/slideshow.js maps the folder URL
to slideshow.html (resolveSiteUrl - used where main.js loads a view and
where attachView re-checks the configured URL) and pushes the image list
into the page from the main process (the page has no fs access), re-read
every minute so added/removed images show up without a restart. Natural
file-name order, crossfade, letterboxed by default; optional
slideshow.json in the folder: interval, shuffle, fit, transition,
recursive. provision_install_app already copies kiosk-app/*.js/*.html,
so new installs and Advanced -> Upgrade pick it up.
Electron: ^42.0.0 -> ^44.0.0. Electron supports the latest three majors
(44/43/42), so 42 drops out when 45 ships. Checked the 43/44 breaking
changes against kiosk-app: nothing it uses changed (no clipboard,
safeStorage, net.request, utilityProcess; BrowserView is deprecated but
still present through 46). The offline bundle builder now retries
Electron's install.js download, which crashes outright if a large
download is cut off mid-stream.
Tested with the real main.js on Electron 44.5.1 under Xvfb + Openbox:
slideshow fills the screen, cycles in order (01, 02, 10) skipping a .txt
file, honours slideshow.json, and an image added mid-run joined the
rotation after the refresh. Offline bundle rebuilt with Electron 44.5.1.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
- CUPS (Addons -> CUPS Printing, from ./install.sh or the Web UI) now
installs from the offline bundle by itself when the kiosk can't reach
the Ubuntu archive: lib/config.sh's run_with_offline_fallback switches
apt to the bundle (kiosk-offline-apt on) around the install and back
afterwards. No-op on manual installs or when online.
- Fix cups_is_installed reporting "not installed" for an installed CUPS:
`dpkg -l | grep -q` under install.sh's pipefail fails when grep exits
early and dpkg takes SIGPIPE. Uses dpkg-query on the one package now.
- Sites: an absolute path (/home/kiosk/docs/menu.pdf) is accepted and
turned into a percent-encoded file:// URL, identically in the Sites
menu (sites_parse_url) and the Web UI (parseUrl, with tests). Electron
42 shows local HTML, PDFs (built-in viewer) and images as-is.
- iso/build-iso.sh checks its build tools (plus the --offline ones) and
offers to `sudo apt install` what's missing; --install-deps does it
without asking. --offline now refuses non-amd64 hosts up front.
Tested: offline chroot of the 26.04.1 server filesystem - action_install_cups
installed CUPS + printer-driver-all/hplip/gutenprint from the bundle with
no network and left apt back on normal sources; webui test suite passes;
Electron 42 rendered a file:// PDF and PNG with spaces/# in the path;
build-iso.sh stopped with the install command when apt-utils was missing
and no TTY, and installed it and built the ISO with --install-deps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
`iso/build-iso.sh --offline` (and the workflow's new `offline` box, on by
default) adds ~420 MB to the ISO and copies it to /opt/kiosk-offline:
- apt/: local repository with every package install.sh's provisioning
and the CUPS addon install ask for, NodeSource's nodejs, and all
dependencies a fresh server install doesn't already have. Resolved
against the source ISO's own package manifest via a synthesized dpkg
status (iso/offline/fake-dpkg-status.py), so apt downloads only what's
missing and makes the same virtual/alternative choices as the target.
- npm/: prebuilt kiosk-app (Electron 42, binary included - it no longer
downloads at npm install time) and Web UI node_modules.
On first boot, kiosk-firstboot installs online as before when there's
internet; with none it waits KIOSK_OFFLINE_WAIT (90s), then runs
`kiosk-offline-apt on` (apt.conf.d drop-in: bundle is the ONLY source)
and turns it off again afterwards. install.sh notices offline mode via
offline_mode_active (lib/config.sh) and installs nodejs from the bundle
and unpacks prebuilt node_modules instead of npm install.
Package lists moved into PROVISION_APT_PACKAGES /
PROVISION_INTEL_APT_PACKAGES / CUPS_APT_PACKAGES so the bundler reads
the same lists install.sh installs from.
Tested: built against the 26.04.1 ISO, then in a network-less chroot of
the ISO's own server filesystem installed every package set, ran
provision_create_kiosk_user/install_nodejs/install_app and
webui_install_app_files in offline mode (Electron runs, no missing
libraries, chrome-sandbox setuid), and ran kiosk-firstboot through the
offline switch with a stub install.sh.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
Manually triggered (workflow_dispatch): pick interactive /
unattended-kiosk / fully-automatic, optionally pre-set the admin user
(password from the KIOSK_ADMIN_PASSWORD repo secret, never a logged
input), hostname and SSH key. Runs iso/build-iso.sh and uploads the ISO
plus its sha256 as a run artifact (uncompressed, 3-day default retention).
build-iso.sh changes to run in CI:
- Password can come from $KIOSK_ADMIN_PASSWORD; with no TTY and no
password it fails clearly instead of hanging on a prompt.
- Stage the repo by fetching the commit rather than `git clone --branch`,
which fails on actions/checkout's detached HEAD; the branch name falls
back to $GITHUB_REF_NAME, then main, and is set to track origin so
Advanced -> Upgrade works on the installed kiosk.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
Found by an emulated install: Subiquity refuses --username admin (a
system group) at boot with 'Username is reserved by the system', after
the user has already written the stick. Check against subiquity's
reserved-usernames list at build time instead.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
iso/build-iso.sh remasters the official 26.04.x live-server ISO (downloaded
and checksum-verified, or --iso PATH) with xorriso, replaying its BIOS/UEFI
boot records unchanged. It adds:
- /autoinstall.yaml: Subiquity autoinstall that answers locale/keyboard/SSH
and by default leaves only network, storage and identity interactive
(--username/--auto-storage/--auto-network/--no-confirm/--fully-automatic
make those automatic too). late-commands copy the repo (a clone of the
current HEAD) to /opt/ubuntu-based-kiosk and enable a first-boot service.
- iso/firstboot/kiosk-firstboot(.service): on first boot, takes tty1, waits
for cloud-init and internet, copies the repo to the admin user's home and
runs ./install.sh as that user with temporary passwordless sudo (removed
on every exit path). Failed/interrupted installs re-run next boot.
install.sh hooks for this:
- KIOSK_UNATTENDED=1 makes ask_yes_no take each prompt's default and skips
the Core Settings menu (--unattended).
- KIOSK_PROVISION_DONE_MARKER is touched by provision_finish before the
reboot prompt, so the caller can tell a finished install from an
interrupted one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cjz8a58eB5P9sUweqd3NHm
The keyboard-type IPC handler set input.value directly and dispatched a
plain "input" event. React (and similar frameworks) patch the value
setter on input/textarea instances to track the last value they set;
setting el.value directly updates that tracker too, so React never
detects a real change and its controlled state stays empty. The next
re-render (moving focus to another field, toggling a checkbox, etc.)
then redraws the input from that stale empty state, wiping out
whatever was typed. Route the writes through the native value setter
instead so the tracker stays out of sync and the dispatched event
actually reaches the framework's handler.
Also give the "Tab" key its own handler that moves focus to the next
focusable element, instead of falling through to the generic branch
that typed the literal word "Tab" into the field.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V1D2cv1KyFoYGnwSdszeXm
Web UI now installs by default during first-time provisioning (fixed
port 8090, no prompt) instead of being opt-in, and can install/
reconfigure CUPS Printing, LMS Server, Squeezelite Player, and Asterisk
Intercom, and check for updates - the addons and Update action asked
for by name.
Privilege model: the web service itself still runs as $KIOSK_USER with
zero ambient sudo. A new narrow, allow-listed root helper
(menus/addon_webui.sh's webui_write_helper_script) is the only way it
ever gains privilege, reachable only via a single-path passwordless
sudo rule generated and validated with `visudo -c -f` before being
installed, and it re-checks its own fixed action allow-list before
dispatching anything. Each allow-listed action is the exact same
interactive action_* function the terminal menu already uses, driven
by piping the right answers on stdin - the same technique this
project's own bash tests already use, so no prompt/mutation refactor
of any addon file was needed. webui/lib/actions.js's stdin sequences
were cross-validated against the real bash functions (not just read),
which caught two real bugs (Squeezelite and Asterisk Intercom both
silently lost their "decline reconfigure" path).
Long-running installs stream live output via Server-Sent Events
(webui/lib/jobs.js), one action at a time.
Full visual redesign: a sidebar shell (Sites/Display/Lockout/Addons/
Update) replacing the single scrolling page, light+dark themes via
prefers-color-scheme, no external font/CDN dependency. Actually driving
the redesigned UI in a headless browser (not just reading the code)
caught a real bug: refreshing an addon's pill/button after a successful
install used to rebuild the whole card, racing (and usually losing to)
the success status/log that job had just written. Fixed to update
pill/buttons in place.
Uninstall-via-web is deliberately still not offered, for any addon.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VfsFSoRqfbRG7XAg5RoE7e
New Addons -> Web UI: a small Node/Express app (webui/) installed as a
systemd service running as $KIOSK_USER, giving a browser-based editor
for Sites & Page Timing, Display & Interaction, and Password Protection
& Lockout - the three Core Settings menus that are pure config.json
read/write with no privileged system mutation involved. Runs with no
sudo at all, since config.json is already owned by $KIOSK_USER.
webui/lib/config.js re-implements lib/config.sh's exact schema and
merge-on-save contract in JS (kiosk-app/main.js already reads the same
config.json directly in JS, so this isn't a new pattern), so it can
never silently clobber fields it doesn't track - the same bug
previously fixed in lib/config.sh's own history.
No login of its own by design: Authelia runs elsewhere, and the
expectation is a reverse proxy (e.g. Caddy) with Authelia forward-auth
in front of it, the same way other self-hosted apps get protected -
Authelia integration is explicitly out of scope for this repo.
Deliberately narrow scope for this first pass: WiFi, Timezone,
Power/Display/Quiet Hours, Complete Uninstall, every other addon, and
everything in Advanced remain terminal-only, since a network-facing
process shouldn't be handed sudo-level system mutation without a lot
more thought than this pass gives it. Wired into Complete Uninstall
(webui_do_uninstall) and Clone Settings (addon-presence detection) the
same way every other addon is.
This is the single-kiosk piece of the web-based GUI this repo's
"Modular Management" notes have mentioned for a while - a central
multi-kiosk fleet dashboard is an intentional follow-up.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VfsFSoRqfbRG7XAg5RoE7e
It never worked reliably in the legacy script, and the modular tool
already covers the same outcome more reliably as two already-tested
pieces run back to back: Complete Uninstall (Core Settings), then
./install.sh again to provision fresh. No code changes - nothing was
ever built for it in the modular tool, this just stops carrying it in
docs/comments as a pending gap.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VfsFSoRqfbRG7XAg5RoE7e
The legacy Upgrade re-extracted main.js/preload.js/etc from its own
heredocs on every run - a mechanism the modular tool has no equivalent
of now that kiosk-app/ and provision/files/ are real files in the git
checkout. The new Advanced -> Upgrade is `git pull` (only after a
clean-tree check, and only as a fast-forward - never an automatic
merge) followed by re-running the same packages/kiosk-app/display/
firewall/power-management steps lib/provision.sh already has for a
fresh install, reused rather than reimplemented. Also offers an
on-demand Electron version check via the existing action_update_electron,
since Electron isn't versioned by this repo.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VfsFSoRqfbRG7XAg5RoE7e
Until now ./install.sh only managed an already-installed kiosk;
ubuntu-based-kiosk.sh was still the only path from a bare Ubuntu Server
box to a running one. install.sh now provisions from scratch too:
packages, kiosk user, Node.js/Electron, LightDM+Openbox autologin,
audio/video/HDMI/power-button hardware setup, and the firewall, then
hands off to the already-migrated Core Settings/Advanced menus for
initial configuration instead of reimplementing that logic again.
- lib/provision.sh: the new provisioning flow, built mostly by calling
existing menus (core_settings_menu, emergency hotspot, virtual
consoles) - cuts it to ~300 lines against the legacy script's
~4,000-line first_time_install().
- lib/electron.sh: electron_install_binary() extracted out of
menus/advanced_electron.sh so provisioning and the existing "Fix
blank screen" action share one implementation.
- kiosk-app/ and provision/files/: the Electron app source and every
system template file, extracted byte-for-byte out of
ubuntu-based-kiosk.sh's heredocs into real files.
- Found and fixed a bash set -e gotcha along the way: testing a
multi-statement function as an if-condition (`if ! some_func; then`)
silently exempts everything inside that function from set -e for the
duration of the call. Fixed in the new provisioning code and in
menus/advanced_electron.sh's pre-existing repair action, which had
the same shape.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VfsFSoRqfbRG7XAg5RoE7e
menus/fleet_profile.sh -> menus/clone_settings.sh. Same functionality
(export/apply portable config.json settings between kiosks); renamed
the file, every function, the menu label, and the on-disk default
profile filename to match. No behavior change.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VfsFSoRqfbRG7XAg5RoE7e
New menus/fleet_profile.sh (Advanced), for the "set up one kiosk, then
stand up a dozen more like it" use case. Not a port of the legacy
Export/Import Settings - a narrower, deliberately-scoped feature:
- Export: writes config.json's portable fields (sites, display/touch/
navigation, lockout, password protection) plus a list of addons
present at export time to a JSON profile file.
- Apply: merges those fields onto a target kiosk's config.json (same
merge-not-replace pattern as save_config, so the target's own fields
survive untouched) and prints a checklist of which listed addons are/
aren't installed on the target.
Deliberately excludes machine-bound credentials rather than silently
mishandling them: Authelia's encrypted password is keyed off
/etc/machine-id and decrypts to garbage elsewhere; a WireGuard private
key is a device identity, reusing one across machines is a peer
conflict; most Asterisk PBXes reject duplicate registrations to the
same extension. Apply prints all three as an explicit "needs a human"
checklist. Non-interactive addon installation (for a fully scriptable
fleet rollout) is a deliberate follow-up, not part of this MVP.
Full command-level stubbed test suite covering export (site/setting
content, Authelia stripped, addon detection) and apply (merge
correctness, target's own Authelia preserved, bad path/invalid JSON
handled cleanly). Full 20-suite regression + real end-to-end menu
navigation via install.sh all pass.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VfsFSoRqfbRG7XAg5RoE7e
New menus/complete_uninstall.sh (Core Settings), the last of the
"destructive trio". Rather than re-implementing every addon's teardown
a second time (the legacy shape), it composes the *_do_uninstall
helpers each addon already has - if an addon's removal logic changes,
Complete Uninstall picks it up automatically.
Every addon menu with an uninstall action (CUPS, VNC, WireGuard,
Tailscale, Netbird, LMS, Squeezelite, Asterisk Intercom) plus
power_schedule's "remove all schedules" and Emergency Hotspot's disable
action were each split into a confirm-and-call wrapper (unchanged from
the user's perspective) and a silent do-the-removal helper that both
the wrapper and Complete Uninstall call.
Bug fix found while composing these: several *_do_uninstall helpers
(CUPS's apt autoremove/apt clean, VNC/WireGuard/Tailscale/Netbird's apt
remove) had a bare, unguarded apt call as their second-to-last
statement. Previously this only risked aborting that one menu action if
the package was already gone. Composed together as sequential calls
inside Complete Uninstall, the same failure would have silently
truncated the entire uninstall sequence partway through. Guarded all of
them with `|| true`.
Non-addon teardown (kiosk user/files, Node.js, LightDM/Openbox,
remaining systemd units/scripts, polkit rules, re-enabling virtual
consoles, final package cleanup) stays inline in
menus/complete_uninstall.sh, since no single addon owns those paths.
Upgrade and Full Reinstall stay in ubuntu-based-kiosk.sh only - both
are coupled to its own heredoc self-extraction of main.js/preload.js/
etc, which has no modular equivalent yet.
Full command-level stubbed test suite exercising the full 12-step
teardown, confirmation-text validation, and reboot prompt. Full
19-suite regression + real end-to-end menu navigation via install.sh
all pass.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VfsFSoRqfbRG7XAg5RoE7e
New in install.sh's Advanced menu, alongside Diagnostics:
- menus/advanced_electron.sh: "Electron Maintenance" - the legacy
"Manual Electron Update" and "Fix Blank Screen" combined into one
submenu, sharing the binary-repair logic (electron_install_binary).
- menus/advanced_factory_reset.sh: "Factory Reset" - wipes config.json
back to defaults only; addons are untouched.
- menus/advanced_virtual_consoles.sh: "Virtual Consoles" - toggles
Ctrl+Alt+F1-F8 terminal login access.
- menus/advanced_emergency_hotspot.sh: "Emergency Hotspot" - auto-starts
a WiFi hotspot if no internet is detected 60 seconds after boot. Its
own runtime script and systemd unit now go through $BIN_DIR/
$SYSTEMD_DIR like every other addon's own files, instead of the
legacy's hardcoded /usr/local/bin and /etc/systemd/system.
That covers 8 of the legacy Advanced menu's 12 entries. Not migrated
this round: Export/Import Settings (pending a decision on rebuilding it
around actual paths vs. a hardcoded step list, or whether the future
web UI replaces the need for it) and Fix Squeezelite Audio (small
enough it may fold into the LMS addon instead of staying standalone).
Full command-level stubbed test suite per file, including set -e safety
checks (declined/failed paths never crash the session) and content
verification for every written file. Full 18-suite regression + real
end-to-end menu navigation via install.sh all pass.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VfsFSoRqfbRG7XAg5RoE7e
New menus/addon_asterisk_intercom.sh, wired into install.sh's Addons
menu. The legacy addon offered Client Only (Baresip SIP client), Server
Only, and Full (server + client), where Server/Full downloaded and ran
a third-party installer from a separate "Easy Asterisk" repository to
stand up a whole Asterisk PBX. That repository has since gone through a
major rework upstream, so this migration drops the PBX-install path
entirely: the addon now only installs Baresip and registers this kiosk
as a SIP extension against an Asterisk server the user already has
running elsewhere. It never installs or manages Asterisk itself. The
legacy script's own three-option version is untouched, same as every
other migrated menu.
Dropped the legacy client path's dependency on the Easy Asterisk repo's
GitHub API for version tracking - now reads the real installed baresip
package version via dpkg instead. Added an uninstall option, which the
legacy addon never had at all.
Bug fix found while testing: an unguarded `ver=$(baresip_installed_version)`
assignment crashed the whole session under set -e the first time status
was checked before Baresip was installed (dpkg-query legitimately fails
when the package isn't there). Guarded with `|| true`.
Full command-level stubbed test suite covering configure (manual/auto-
answer, TLS port bump, apt-install failure) and uninstall (keep/purge
config) for both fresh and already-configured states. Full 13-suite
regression + real end-to-end menu navigation via install.sh all pass.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VfsFSoRqfbRG7XAg5RoE7e
New menus/addon_lms_squeezelite.sh: install/reconfigure/uninstall for an
LMS (Lyrion/Logitech Media Server) server and a Squeezelite player,
wired into install.sh's Addons menu. Squeezelite's own start script and
systemd unit now go through $BIN_DIR/$SYSTEMD_DIR like every other
addon instead of hardcoded /usr/local/bin and /etc/systemd/system; LMS's
own apt repo/GPG key/ufw rules stay at their real fixed system paths,
same approach as CUPS.
Fixed a real unguarded-pipeline bug from the legacy install_lms():
`sudo systemctl enable "$service_name" 2>&1 | tee ...` made the exit
status depend on tee (always 0) instead of systemctl enable, silently
swallowing real enable/start failures. Now uses enable_and_start_units().
Fixed is_service_enabled() (shared helper, backported into the legacy
script too): its list-unit-files pre-check never matched a bare service
name, so it always fell through to "not enabled" regardless of the real
state. Dropped the dead pre-check.
Full command-level stubbed test suite covering install/reconfigure/
uninstall for both LMS and Squeezelite, including the repo-vs-fallback-
download path, undetectable-service-name path, and enable/start-failure
path. Full 12-suite regression + real end-to-end menu navigation via
install.sh all pass.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VfsFSoRqfbRG7XAg5RoE7e
Third and biggest Addon migrated: menus/addon_remote_access.sh - VNC
(x11vnc), WireGuard, Tailscale, and Netbird, each with its own install/
connect/status/uninstall flow. Same risk class as CUPS (real apt
packages, real system state) but broader in scope: Tailscale and
Netbird install via the vendors' own documented `curl -fsSL <url> | sh`
method, preserved exactly as-is rather than redesigned.
- lib/config.sh: new $WIREGUARD_DIR, same pattern as $SYSTEMD_DIR/
$BIN_DIR/etc - nothing in this file hardcodes /etc/wireguard.
- lib/menu.sh: promoted power_schedule.sh's enable_and_start_timers()
to a shared enable_and_start_units() (works for services now too, not
just timers) - Remote Access needed the identical enable+start-with-
graceful-failure-reporting pattern for x11vnc and wg-quick@, so this
is fixed once and reused rather than duplicated a second time.
power_schedule.sh's four call sites renamed to match.
Found and fixed a real framework-level bug while building this file:
run_menu()'s *handler* call has been `|| true`-guarded since v2.1.0,
but the *status function* call (`"$status_func"` on its own line) was
still completely bare. A status function's entire job is read-only
display, but if it contains so much as a pipeline whose grep matches
nothing - which pipefail turns into a pipeline failure even though the
actual last command in it (e.g. sed) succeeds - that bare call would
crash the *entire session*, not just fail to show status text. Found
while writing wireguard_status()'s `sudo wg show | grep ... | sed ...`
and deliberately verifying its exact failure mode rather than assuming
run_menu already covered it. Fixed once in run_menu() itself
(lib/menu.sh), protecting every status function across every menu -
present and future - the same "fix once at the framework level"
pattern as the v2.1.0 handler fix.
Given the framework fix meant this class of bug had been silently
possible since v2.1.0, audited every existing status function across
every already-migrated menu for the same specific shape (a bare
`var=$(...)` assignment from a grep-based pipeline, not embedded in an
echo and not already guarded - embedded substitutions and if-condition
contexts are both already safe on their own). Found and fixed one real
instance in power_schedule_status(). menus/addon_remote_access.sh's own
two equivalent pipelines (wireguard_status, netbird_status) were
written with `|| true` from the start once the pattern was identified.
Verified:
- New scratch/stub test for addon_remote_access.sh, with curl stubbed
separately from sudo (Tailscale/Netbird's install scripts must never
reach the real network regardless of what sudo intercepts) and a
belt-and-suspenders `sh` stub in case anything got past curl: full
status/menu-builder coverage for all four sub-areas in their real,
unstubbed "not installed" state (none of the four tools exist in this
sandbox); VNC install/change-password/uninstall with systemd unit
content verified (correct $KIOSK_USER/$KIOSK_HOME substitution);
WireGuard install, paste-config (content written correctly to scratch
$WIREGUARD_DIR), and uninstall - including documenting a genuine
cat-until-EOF test-harness limitation (a redirected pipe's EOF is
permanent for the whole stream, unlike a real terminal's per-read
Ctrl+D, so only the config's *default* name is testable through
simple stdin redirection - inherent to the design, matches the legacy
script's identical `cat`-based approach, not a bug); Tailscale and
Netbird install/connect-interactive/connect-with-key/uninstall; and
all four cancel paths confirmed to make zero sudo calls.
- Full regression: re-ran all 11 prior scratch/stub suites after the
lib/menu.sh and power_schedule.sh changes - all still clean.
- End-to-end: ran the real install.sh as a genuine non-root, non-
"kiosk" user through Addons -> Remote Access -> all four sub-menus in
turn, each showing accurate real (unstubbed) "not installed" status,
selecting Install, declining the confirmation, and returning cleanly
- zero invalid-choice errors, clean exit code 0.
Per user decision: backport just the config-clobbering fix from v2.6.0
(lib/config.sh) into the legacy single-file installer's own
save_config(), independent of migrating the rest of that menu into
./install.sh.
The bug: save_config() rebuilt config.json from a fixed list of known
fields via `jq -n`, silently deleting anything it didn't know about -
specifically autheliaURL/autheliaUsername/autheliaEncryptedPassword,
written by configure_authelia()'s own careful `. + {...}` merge.
Configuring Authelia and then visiting Sites, Touch Controls,
Navigation, or Password Protection (all of which call save_config)
silently deleted the Authelia credentials. Real, currently-shipping
credential-loss bug, unrelated to whether the rest of that menu is ever
migrated - didn't need to wait for a full pass.
Fixed the same way as lib/config.sh: merge the known fields onto
whatever's already in config.json (`. + {...}`) instead of rebuilding
from nothing, with a `jq empty` validity check falling back to `{}` if
the existing file is missing or corrupt. This is a standalone fix to
one function only - nothing else about Sites/Touch/Navigation/Authelia
changed, and none of that is migrated by this commit.
Verified before touching the shipping copy: extracted the exact
save_config() function (now lines 3682-3805) into an isolated test
harness with stubbed dependencies (kiosk_user_exists, is_service_active,
log_success/warning), seeded a stub config.json with Authelia-style
fields via the same `. + {...}` merge configure_authelia() uses, called
save_config() a second time simulating a visit to an unrelated menu,
and confirmed the Authelia fields survive while an actual settings
change (duration 60 -> 90) still correctly takes effect. Also verified
the corrupt-JSON and missing-file edge cases don't crash the function.
Full syntax check on the whole 12,000+ line script, and the entire
modular test suite (11 scratch/stub suites), both still clean.
Second Addon migrated: menus/addon_authelia.sh (encrypted SSO
credentials, AES-256-CBC with a key derived from /etc/machine-id via
scrypt - same algorithm main.js decrypts with - plus the full
Dockerized server-side setup instructions, now viewable again later
without reconfiguring).
Investigating how to wire its three config.json fields (autheliaURL/
autheliaUsername/autheliaEncryptedPassword) into lib/config.sh surfaced
a real, currently-shipping bug that has nothing to do with Authelia
specifically: save_config() did a full `jq -n` rebuild of config.json
from a fixed list of known fields - identical to what the legacy
script's own save_config still does. The legacy configure_authelia()
writes its three fields via a careful `. + {...}` merge that preserves
everything else already in the file, but neither save_config knew those
fields existed - so the next time a user visited Sites, Touch Controls,
Navigation, or Password Protection (all of which call save_config),
their Authelia credentials were silently deleted. This bug already
existed in the shipped single-file installer; it was ported faithfully
into lib/config.sh's first version because no test happened to set an
untracked field before calling save_config.
Fixed in lib/config.sh: save_config now merges its known fields onto
whatever's already in config.json (jq `. + {...}`) instead of rebuilding
the file from nothing, with a `jq empty` validity check falling back to
`{}` if the existing file is missing or corrupt. Any field this tool
doesn't track - Authelia's three today, anything else a future addon
adds tomorrow - now survives automatically. autheliaURL/
autheliaUsername/autheliaEncryptedPassword are also tracked fields in
their own right now (load_existing_config/save_config), consistent with
every other config.json field this tool manages, giving Authelia both a
direct fix and the general safety net.
The equivalent bug still exists, unfixed, in ubuntu-based-kiosk.sh's own
save_config - noted in both that script's changelog and the Readme's
"Modular Management" section as an open question: whether to backport
just that one fix into the legacy script now, independent of the wider
migration, given it's a real credential-loss bug affecting the
currently-shipping installer today.
Verified:
- New dedicated test (test_save_merge.sh) proving the save_config fix
itself: seeded config.json with a simulated untracked field via the
same `. + {...}` merge Authelia's own code uses, called save_config
from an unrelated context (Sites deleting a tab), and confirmed the
untracked field survived while the tab deletion still correctly took
effect (not undone by the merge) - plus corrupt-JSON and
missing-file edge cases both handled without crashing.
- New scratch-config test for addon_authelia.sh using REAL encryption
(this sandbox has both Node and /etc/machine-id): configured with a
real password, then decrypted the stored ciphertext using main.js's
exact algorithm (independently reproduced in the test) and confirmed
it recovers the original password exactly - true interoperability,
not just "some ciphertext was produced." Also covered cancel paths,
clearing the configuration, the encryption-unavailable failure path,
and confirmed Authelia's config survives an unrelated Sites save.
- Full regression: re-ran all 9 prior scratch/stub test suites after
both the lib/config.sh changes - all still clean.
- End-to-end: ran the real install.sh as a genuine non-root, non-
"kiosk" user with a seeded minimal config.json, through Addons ->
Authelia -> Configure with a real URL/username/password -> confirmed
the resulting config.json on disk, and independently decrypted the
stored password for real using main.js's algorithm to confirm it
matches exactly. Clean exit code 0 throughout.
First Addon migrated: menus/addon_cups.sh (install, reconfigure for
network access, complete uninstall/purge). Different risk profile from
everything migrated so far - it genuinely mutates real system state
(apt install/remove --purge, /etc/cups, ufw) at fixed paths CUPS itself
doesn't let us relocate, unlike the systemd/cron/bin paths this project
already controls via $SYSTEMD_DIR etc. Only the polkit rule's directory
is parameterized ($POLKIT_DIR, lib/config.sh, since that one is ours to
place); everything else gets full command-level `sudo` stubbing in
every test - there is no scratch equivalent for a real apt-managed
subsystem's own file layout. Also added $BUILD_USER (the admin account
actually running the tool, as opposed to $KIOSK_USER) since CUPS needs
to grant it lpadmin group membership.
Restructured install.sh's top-level menu into Core Settings / Addons /
Advanced (matching the legacy tool) instead of one flat list, now that
Addons exists as its own category - cheap to do with one item in it,
much more annoying to retrofit once the flat list has fifteen.
Two bugs caught and fixed before they shipped:
- A "wait for CUPS to start" retry loop used a bare `cmd1 && cmd2 &&
break` as its body while "simplifying" the legacy script's `if cmd1
&& cmd2; then break; fi`. Being inside a loop doesn't protect a bare
&&/|| list from set -e - only if/while/until conditions and the
protected side of &&/|| do that - so the first command failing on an
early iteration (near-certain right after a fresh install, before
CUPS has actually started) would have crashed the entire session.
Restored the `if` form; noted the lesson in the file's own header
comment since it's a general trap, not CUPS-specific.
- Resolved real uncertainty, rather than assuming: how far does
run_menu's `handler || true` guard (v2.1.0) actually protect? Wrote a
minimal isolated test (a bare `false` three function calls deep,
called via `outer || true` at the top) and confirmed bash's errexit
exemption for the left side of `||` covers the *entire* evaluation,
arbitrarily deep through function calls - not just the immediately
invoked function. So the session-crash risk this project has been
chasing since v2.1.0 is already covered end-to-end by that one fix.
Per-statement guards (`|| true`, explicit `if`) still earn their keep
for a different reason: without them a deep failure bubbles silently
past the menu actually responsible for it to wherever the nearest
`|| true` happens to sit, which can be several menu levels above
where the user actually was - not a crash, but a confusing jump.
Verified:
- Full regression: re-ran every existing scratch-config/stub test suite
after the lib/config.sh change (new $BUILD_USER/$POLKIT_DIR) and
after the install.sh restructuring - all still clean.
- New scratch/stub test for addon_cups.sh: full state-machine coverage
(not installed -> decline -> install -> running -> reconfigure ->
stopped -> start -> uninstall decline -> uninstall confirm -> not
installed again) with every `sudo` call intercepted and only `rm`
targeting the scratch $POLKIT_DIR ever actually executed; confirmed
the polkit rule's content and that declining install makes zero sudo
calls. Added both apt-failure paths (update fails, install fails)
and confirmed the tool reports clearly and returns to the menu
instead of dying, exercising the exact bug class just fixed.
- End-to-end: ran the real install.sh as a genuine non-root, non-
"kiosk" user through the full new three-level structure - Core
Settings -> Sites -> back -> back, Addons -> CUPS -> declined install
(using this container's real, unstubbed dpkg check, correctly
reporting "not installed" and making no apt/systemctl calls) -> back
-> back, Advanced -> Diagnostics -> System status -> back -> back ->
Exit. Zero invalid-choice errors, clean exit code 0 throughout.
Deliberately skipped Upgrade/Full Reinstall/Complete Uninstall for now:
all three are large (130-250 lines), genuinely destructive (wipe/
reinstall the kiosk), and Upgrade specifically is coupled to the legacy
script's own self-extraction mechanism (it greps its own running source
for embedded heredocs to pull out main.js/preload.js) - there's no
modular equivalent to migrate it to yet, since those files don't exist
as separate assets outside the monolith. Migrated Diagnostics instead:
4 of the legacy Advanced menu's 12 items (System Status, View Logs,
Audio Diagnostics, Network Test), all read-only except one optional
"play a test sound?" prompt - a deliberate change of pace with no
destructive-action risk to design around, after Sites/WiFi/Power.
- lib/menu.sh: ported get_vpn_ips alongside get_ip_address.
- menus/diagnostics.sh: straight port, using $KIOSK_USER/$KIOSK_HOME
throughout instead of the legacy code's mix of the variable and a
hardcoded "kiosk" literal.
Bug fixed, same set -e-safety class as v2.1.0's run_menu fix and
v2.3.0's netplan/systemctl fixes, but a bigger batch this time: nearly
every diagnostic command here was a bare unguarded statement whose
*expected, common* failure - no lightdm running, no audio hardware, no
network, missing log files, ping/nslookup not even installed - would
have crashed the entire session instead of reporting "not found" and
continuing. A diagnostics tool has to be the most crash-proof code in
the project, since it exists to run when something is already broken.
Fixed at every call site: systemctl status | head, tail on lightdm's
log, journalctl, ping, nslookup, and three pactl-backed variable
assignments.
Also noted for future menus in this migration: writing `local var;` and
`var=$(cmd)` as separate statements (good practice, and how several
earlier real bugs were caught) removes an accidental safety net -
`local x=$(cmd)` on one line masks the substitution's exit code with
`local`'s own always-success status. Splitting them is correct, but
each split assignment needs an explicit `|| true` (or real fallback)
where failure is expected and non-fatal, rather than relying on that
masking by accident. Caught three instances of exactly this while
writing this file fresh, not just porting old bugs.
Verified:
- Full regression: re-ran every existing scratch-config/stub test suite
(sites, display, timezone/pin, lockout, power schedule + RTC, wifi) -
all still clean after the lib/menu.sh change.
- New test for diagnostics.sh, exercised mostly for real (no
destructive-mutation risk here, so minimal stubbing needed): system
status, all three log views (including the "no such file" paths for
lightdm log and electron log), full 8-step audio diagnostic with test
sound declined, and network test - all report gracefully instead of
crashing, confirmed by re-running after each fix until every bare
unguarded statement was accounted for.
- End-to-end: ran the real install.sh as a genuine non-root, non-
"kiosk" user, navigating Diagnostics -> System status -> View Logs ->
System journal -> Audio diagnostics (declined test sound) -> Network
test -> exit. Confirmed every diagnostic path completes and returns
to its menu cleanly (exit code 0) even with ping/nslookup missing and
no audio hardware/network present in this environment.
By far the riskiest menus migrated so far. Both can affect real system
state outside config.json in ways that are hard to reverse: WiFi
rewrites live netplan config and, over SSH, can disconnect the very
session configuring it; power scheduling can shut the physical machine
down and wake it via RTC.
- lib/config.sh: new $SYSTEMD_DIR/$CRON_D_DIR/$BIN_DIR/$NETPLAN_DIR,
same `: "${VAR:=default}"` pattern as $KIOSK_DIR. Nothing under
menus/ hardcodes /etc/systemd/system, /etc/cron.d, /usr/local/bin, or
/etc/netplan directly, so every test in this change points them at
scratch space instead of ever touching this sandbox's real systemd
units, cron, or network config.
- lib/menu.sh: ported get_ip_address (also fixing its "No IP" fallback,
which never actually fired before - `hostname -I | awk` always exits
0 even on empty output).
- menus/wifi.sh: apply_wifi_config split out from wifi_menu specifically
so tests can drive the netplan-writing logic without needing real
scan hardware. Preserves the legacy netplan backup, 60s SSH watchdog,
and restore-on-failure behavior exactly.
- menus/power_schedule.sh: power schedule (+ RTC wake), display
schedule, quiet hours, and an Electron reload timer (with its own
nested run_menu, mirroring the legacy configured/not-configured
dispatch), plus remove-all. Deliberately excludes the legacy
dispatcher's "Test schedules & system" - a shared diagnostics submenu
(audio/network/keyboard tests) that isn't specific to scheduling and
belongs with a future Advanced/Diagnostics migration instead.
Bugs found and fixed along the way, none papered over:
- The legacy dispatcher refused to open "Configure power schedule" at
all without RTC hardware, even though shutdown-only mode never needed
RTC. Now always available.
- None of the six HH:MM prompts across these menus (shutdown, wake,
display off/on, quiet start/end, custom Electron reload time) were
validated before - plain `read`, no format check. All now go through
ask_time.
- set -e safety (same class as the v2.1.0 run_menu fix), three more
instances: `ls *.yaml` when no netplan file exists still fails under
pipefail even with stderr silenced (masked in practice by cloud-init
usually leaving a file behind); the restore-and-reapply `netplan
apply` after an initial failure was a bare unguarded statement; and
`systemctl enable`/`start` after writing each of the four timer pairs
was unguarded too - caught only by testing in an environment without
a live systemd, but a real enable/start failure on actual hardware
(bad unit, daemon-reload skipped, ...) would hit the exact same crash.
Added a shared enable_and_start_timers() helper used at all four call
sites; all now report a clear warning and return to the menu instead
of taking the session down.
Testing discipline for this round, given the risk:
- No automated test calls the real netplan/nmcli/iw/wpa_cli/systemctl -
confirmed no WiFi tools or `wl*` interface exist in this sandbox, so
wifi_menu's own tools-check safely short-circuits before touching
anything; apply_wifi_config's actual YAML/backup/failure-recovery
logic is tested with sudo/netplan/get_ip_address stubbed instead.
- One stubbing pitfall caught and fixed in the test itself: `nohup sudo
bash "$watchdog" ... &` execs nohup as a real external binary, which
then execs the real sudo - a bash function stub named `sudo` does NOT
intercept that, only stubbing `nohup` itself does. Verified via pgrep
that no real watchdog process or `sleep 60` was ever spawned.
- power_schedule.sh tested with SYSTEMD_DIR/CRON_D_DIR/BIN_DIR pointed
at scratch dirs and only `sudo systemctl` stubbed (tee/rm/chmod/cp
left real, since they only ever touch scratch paths): full lifecycle
for all four schedule types plus remove-all, the RTC-available branch
(including the wake-time-before-shutdown-time hour/day wraparound
arithmetic) via a stubbed rtc_wake_available, and the new
enable_and_start_timers failure path via a stub that fails `enable`
specifically.
- End-to-end: ran the real install.sh as a genuine non-root, non-
"kiosk" user for both menus. WiFi correctly short-circuits on missing
tools without crashing. Power/Display/Quiet Hours (SYSTEMD_DIR/
CRON_D_DIR/BIN_DIR redirected to scratch space) configured all four
schedule types in sequence including the nested Electron Reload menu,
survived four consecutive real "systemctl enable/start failed"
warnings (this container has no live systemd) without the session
dying, then removed everything - confirmed the scratch dirs ended up
empty and config.json was never touched (correctly out of scope for
this menu).
Fifth menu migrated onto lib/menu.sh + lib/config.sh: menus/lockout.sh
covers enable/disable, changing the password, inactivity timeout, daily
lock time, and boot password. The password is SHA-256 hashed before
it's ever assigned to LOCKOUT_PASSWORD (matching main.js's comparison
logic) - verified by test that the stored value is the correct hash and
never plaintext.
Rewrote the legacy configure_password_protection's linear "ask
everything, confirm save at the end" wizard as the same immediate-save
pattern used by every other migrated menu: each action (change
password, change timeout, toggle boot password, ...) is a complete,
standalone change, consistent with Sites/Display/Timezone/Hidden PIN.
LOCKOUT_ACTIVE_START/END are deliberately left untouched - per the
Readme they're inert leftover fields the app ignores, so lib/config.sh
just carries whatever is already in config.json through unchanged.
Testing this menu surfaced a real gap before it ever shipped: lib/menu.sh
never had ask_time/validate_time at all (only validate_integer/ask_integer,
ask_url, etc were ported when the framework was first built) - "set a
daily lock time" would have failed for every single user with
"ask_time: command not found". Ported both from the legacy script.
Also promoted the ON/OFF toggle-label helper (previously private to
menus/display.sh as display_onoff) to a shared onoff() in lib/menu.sh,
since menus/lockout.sh needed the same thing and menu files should only
ever depend on lib/, never on each other.
Bumped SCRIPT_VERSION to 2.2.0 with matching changelog entries in the
script header and Readme, and updated "Modular Management" to list the
new menu and drop Password Protection & Lockout from the "not yet
migrated" list.
Verified:
- Full regression: re-ran the Sites, Display, Timezone/PIN scratch-config
suites after every change in this round (the onoff refactor, and again
after adding ask_time) - all still clean.
- New scratch-config test for lockout.sh: enable (password+timeout+daily
lock+boot toggle), independently recomputed the expected SHA-256 hash
and confirmed it matches config.json exactly, change password, change
timeout, clear daily lock, toggle boot password, disable (confirmed
every field clears), and that the menu builder's options correctly
differ between the enabled and disabled states.
- End-to-end: ran the real install.sh as a genuine non-root, non-"kiosk"
user - Lockout menu -> enable protection with a real password entered
via the masked prompt -> set 20m timeout, 23:00 daily lock, boot
password on -> confirmed the menu redraws with the new state -> clean
exit (code 0). Checked the resulting config.json and file permissions
on disk.
Two more menus migrated onto lib/menu.sh + lib/config.sh, chosen
specifically because neither touches config.json - a third and fourth
shape for the framework (a system command via timedatectl, and a flat
PIN file), on top of Sites' list CRUD and Display's JSON toggles.
- menus/timezone.sh: also replaces the legacy script's hand-numbered
18-entry case statement with a plain data list (TIMEZONE_COMMON_ZONES)
plus one handler that reads the number run_menu hands it - adding or
removing a zone never touches numbering anywhere else. Required a
small run_menu addition: handlers now receive the chosen 1-based
number as $1, so one handler can serve a whole data-driven list
instead of needing a wrapper function per entry.
- menus/hidden_pin.sh: set/disable/reset the PIN gating hidden pages.
Testing menus/timezone.sh surfaced a real bug before it ever shipped:
this whole tool runs under `set -e`, and set_timezone() rejecting an
invalid zone via a bare `return 1` as its last statement took down the
*entire* install.sh session, not just that one action - a single typo
would silently drop the user back to their shell. Fixed at the
framework level in lib/menu.sh (run_menu now absorbs a failed handler's
exit code) rather than patching set_timezone alone, since any future
menu could hit the same trap. Verified against the real install.sh as a
genuine non-root user: an invalid timezone now logs an error and
redraws the Timezone menu instead of killing the session (confirmed
exit code 0 at the end of the run). Note this specific hazard was
introduced by this session's own return-1 idiom, not inherited from the
legacy script, which never uses a bare return 1 in these functions.
Also per the user: left the old configure_sites/configure_touch_controls/
configure_navigation_security/configure_optional_features functions in
ubuntu-based-kiosk.sh untouched for now (still carrying the v2.0.0
settings-clobber and reorder bugs) rather than removing them - they'll
be retired in one pass once enough of Core Settings/Addons/Advanced is
migrated. Bumped SCRIPT_VERSION to 2.1.0 with matching changelog entries
in the script header and Readme, and updated the Readme's "Modular
Management" section to state plainly what is and isn't migrated yet.
Verified:
- Full regression: re-ran the Sites and Display scratch-config test
suites against the updated run_menu signature - both still clean.
- New scratch-config tests for hidden_pin.sh (set/mismatch/reject/
disable/reset, correct file permissions) and timezone.sh (builder
entry count, common-zone pick by index, manual entry with legacy
US/* alias normalization, region search + cancel, invalid-zone
rejection) - all correct, with timedatectl/sudo stubbed only where
needed to avoid mutating this sandbox's real system clock/timezone.
- End-to-end: ran the real install.sh as a genuine non-root, non-"kiosk"
user, navigating Timezone -> manual entry -> invalid zone -> confirmed
no crash and a normal return to the menu, then Hidden Site PIN -> set
a PIN -> confirmed the file on disk (mode 600, correct content) ->
clean exit (code 0).
- menus/display.sh: second menu migrated onto lib/menu.sh + lib/config.sh,
covering touch gesture mode, link navigation security, and the
pause/keyboard/navigation button toggles (previously three separate
Core Settings entries). Deliberately a different shape from Sites
(toggle list vs. list CRUD) to exercise the framework more broadly.
Wired into install.sh's top-level menu alongside Sites.
- Renamed ubuntu-based-kiosk-v1.0.3.sh -> ubuntu-based-kiosk.sh so the
installer can be updated in place instead of growing a new
version-numbered filename every release; released versions are now
tracked via git history and the in-script changelog. Updated all
Readme download/re-run commands accordingly. Older versioned files
(ubuntu-based-kiosk-v*.sh, install_kiosk_*.sh) are left in place as
archived releases.
- Bumped SCRIPT_VERSION to 2.0.0 (new script-level changelog entry) and
the Readme version/changelog to match, given the new modular
management path, the rename, and the two real bugs fixed along the
way (settings clobbered on save, off-by-one in reorder).
Verified before moving on to the web admin work:
- Regression: re-ran the full Sites scratch-config test suite (add,
edit, delete, reorder, home) - still clean, no invalid-input paths hit.
- New: scratch-config test for every display.sh action (touch mode,
navigation security, all three toggles), confirming values persist
through save/reload and that a previously-added site survives
untouched across Display-menu saves.
- End-to-end: ran the real install.sh (not just sourced functions) as a
genuine non-root, non-"kiosk" user with real sudo, driving actual menu
input through Sites -> add a page -> Display -> toggle a setting ->
exit. Confirmed final config.json on disk matches every action taken,
and both guard clauses (run as root; no installed kiosk found) fire
correctly.
Start of pulling the menu system out of the 12k-line single-file
installer so individual menus can change without risking the rest of
the script (network, VNC, addons, etc). This is groundwork for the
planned web-based management UI, which will share the same lib/config.sh
read/write layer instead of duplicating it.
- lib/menu.sh: generic numbered-menu framework (auto-numbered entries,
"0" always exits/returns) plus the validated input helpers menus need.
- lib/config.sh: single load/save for config.json. Fixes a latent bug
where the old Sites menu wrote config.json without first loading
swipe/navigation/lockout settings, silently resetting them to
defaults on save.
- menus/sites.sh: Sites & Page Timing fully migrated - add/edit/delete/
reorder pages, set duration (auto-rotate/manual/hidden) and home page.
Also fixes an off-by-one in the ported reorder logic (moving an item
landed one slot short of the requested position) caught by testing.
- install.sh: new entry point for managing an already-installed kiosk
via `git clone` + `./install.sh`, wired to the Sites menu. Does not
yet replace first-time provisioning, which still uses the existing
single-file installer.
All new site CRUD/reorder/home-page paths were exercised against a
scratch config.json (add with/without basic auth, edit duration, set
home + timeout, 2- and 3-item reorders in both directions, delete) to
confirm the resulting config.json matches expectations.
Previously nothing switched PipeWire's default sink, so audio always stayed on the laptop/built-in speakers regardless of HDMI mirroring. Adds /usr/local/bin/kiosk-audio-route.sh, which finds a sink whose name contains 'hdmi' and makes it the default (moving any already-playing streams onto it) whenever an external display is connected, falling back to the non-HDMI sink when it isn't. Called from autostart once PipeWire is confirmed ready, and from kiosk-hotplug.sh alongside the existing display mirroring on every plug/unplug event. Documents the behavior and manual fallback in the README.
Previously mirroring used 'xrandr --auto', which picks the external output's own native resolution (e.g. a TV's 1920x1080) rather than matching the kiosk/laptop panel's resolution. Extracts the shared mirroring logic (previously duplicated between autostart and the hotplug handler) into /usr/local/bin/kiosk-mirror-display.sh, which now checks whether the external output natively lists the primary's resolution and uses it directly, or generates a matching mode with cvt and forces it via --newmode/--addmode when it doesn't. Documents the behavior and a manual fallback for displays that reject non-native CVT timings in the README.
Adds a udev rule (DRM 'change' events) that triggers a new kiosk-hotplug.service, which re-runs the same xrandr mirroring logic as the Openbox autostart script. External displays plugged in after boot now get mirrored without requiring a login or lightdm restart. Named kiosk-hotplug (not kiosk-display-*) to avoid collision with the existing kiosk-display-* wildcard cleanup in remove_all_schedules(). Also removes the udev rule during Complete Uninstall and documents the behavior in the README.
Openbox autostart now detects any connected output beyond the primary and mirrors the kiosk content onto it via xrandr, instead of leaving detected-but-unused external displays dark. Also documents the behavior, its limitation (applies at session start, not live hotplug), and adds an HDMI troubleshooting section to the README.
apt install already includes curl and git; adds net-tools and ncdu to the same idempotent apt install so they're present on fresh installs without duplicating already-installed packages.
Clarifies that the installer script can be downloaded on another machine and transferred via USB, while noting the kiosk machine still needs internet access during install for apt/npm packages.
Two bugs caused the login screen to appear on new Ubuntu 24.04 server hardware:
1. Ubuntu 24.04's PAM config checks 'user ingroup nopasswdlogin', not 'autologin'.
Add kiosk user to nopasswdlogin group (and autologin for older versions).
2. The upgrade path never wrote /etc/lightdm/lightdm.conf.d/10-kiosk.conf,
so on new hardware running through upgrade the file simply didn't exist.
Refactor: extract configure_lightdm_autologin() shared helper called from
both fresh install (step 19/27) and upgrade, so both paths are consistent.
Also use [Seat:*] instead of [SeatDefaults] for forward compatibility.
https://claude.ai/code/session_01VQ13Fwq4MXxwThLfCXBeGr
[SeatDefaults] is silently ignored by LightDM on newer Ubuntu versions.
Replace with [Seat:*] which is the correct section name for Ubuntu 22.04+.
Also add the kiosk user to the autologin group, which newer Ubuntu
requires for passwordless autologin to work.
Without these fixes, LightDM shows the login screen instead of
auto-logging in and launching the kiosk app.
https://claude.ai/code/session_01VQ13Fwq4MXxwThLfCXBeGr
/home/kiosk is mode 700 so root cannot traverse it. Every [[ -f ]] or
[[ ! -f ]] check on paths inside /home/kiosk was silently returning
'not found' even after the kiosk user had successfully written the file.
Replace all three [[ ! -f "$electron_bin" ]] checks and the
[[ -f "$sandbox" ]] check in install_electron_binary with
sudo -u "$KIOSK_USER" test -f so they run in the kiosk user's
security context and can actually see the files.
https://claude.ai/code/session_01VQ13Fwq4MXxwThLfCXBeGr
Accidentally dropped this line when rewriting the extraction block.
mktemp creates the file as root:root 600, so sudo -u kiosk unzip
gets 'Permission denied' trying to open the zipfile.
https://claude.ai/code/session_01VQ13Fwq4MXxwThLfCXBeGr
The previous fix incorrectly ran unzip as root, which fails because
/home/kiosk is not accessible to root. The kiosk user is the right
actor for the extraction, but two things blocked it:
1. node_modules/electron/dist/ can be owned by root when npm's electron
postinstall runs with --unsafe-perm, so the kiosk user gets
'Permission denied' trying to write there. Fix: sudo chown -R the
electron directory to the kiosk user before extracting.
2. With set -euo pipefail active (upgrade call had no || guard), a
failed unzip or chmod would abort the script silently before the
diagnostic error messages could print. Fix: add || true to both
commands so the function always reaches the explicit -f check which
prints the real error and returns 1. The upgrade call already has
|| { log_error ...; return 1; } from the previous commit.
https://claude.ai/code/session_01VQ13Fwq4MXxwThLfCXBeGr
Two bugs combined to cause the 'Electron binary download failed' error
even though the zip downloaded and unzip reported inflating all files:
1. The upgrade path called install_electron_binary bare (no ||), so
set -euo pipefail was active inside the function. Any failing command
(e.g. chmod on a file that wasn't written) killed the script before
the error messages printed. Fresh install used || exit 1, which
disables set -e inside the function body. Upgrade now uses
|| { log_error ...; return 1; } to match.
2. The unzip ran as the kiosk user, but node_modules/electron/dist/ can
be owned by root when npm's electron postinstall script runs with
--unsafe-perm. The kiosk user can't write there, so unzip's write
errors go to stderr (not visible in the log) while inflating: lines
still appear on stdout. The binary is never actually written.
Fix: run mkdir/unzip/chmod as root, then chown -R to kiosk.
https://claude.ai/code/session_01VQ13Fwq4MXxwThLfCXBeGr
mktemp creates the tmp zip owned by root with mode 600.
sudo -u kiosk unzip then fails with "Permission denied".
Add chmod 644 immediately after download so the kiosk user can read it.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
When the script is run via curl|bash or wget|bash, BASH_SOURCE[0] is a
pipe descriptor, not a real file. The upgrade function grep-extracts
heredocs from the script file, so it fails with a confusing path error.
Fixes:
- Set SCRIPT_FILE global at startup (empty string when piped)
- upgrade_kiosk() checks SCRIPT_FILE before asking "Continue?" and shows
a clear message explaining how to download the script to a file first
- Removes the silent failure path (no more cryptic "Cannot find script at
/proc/.../pipe:[...]" error)
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
Fresh Ubuntu 24.04 minimal installs don't include unzip. The wget fallback
in install_electron_binary() downloaded the 120MB Electron zip successfully
but then failed on the unzip call. Two fixes:
1. Add unzip to the main apt install step so it's always present.
2. Auto-install unzip inside install_electron_binary() as a safety net for
upgrades on existing systems that may not have it.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
Drop the MatchProduct "Finger" restriction from the xorg libinput rule,
leaving only MatchIsTouchscreen "on". MatchIsTouchscreen is set by udev
from hardware capabilities, so it matches finger touch screens of any
brand (ELAN, Goodix, eGalax, Wacom, etc.) while never matching keyboards,
mice, or pen/stylus digitizers (which are tagged as tablets, not
touchscreens). This makes the script work on any touch hardware without
hardcoding device names. Behavior on existing Wacom machines is unchanged
since their finger device matched either way.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
Two genuinely separate root causes were behind the dead touchscreen:
1. GNOME keyring grab — under LightDM autologin the keyring stays locked.
When Chromium accessed it, the gcr-prompter unlock dialog grabbed all
keyboard and touch input at the X level. The app rendered (timers ran)
but ignored every tap and keypress. Fix: --password-store=basic stops
Electron from using the keyring, so the dialog never appears.
2. Wacom driver single-touch emulation — the wacom X driver only does
single-touch pointer emulation and never passes real multitouch to
Chromium, so 1-finger and 2-finger swipe gestures could not fire.
Fix: force the finger touch device to the libinput driver via
/etc/X11/xorg.conf.d/99-finger-libinput.conf. libinput delivers proper
XI2 multitouch which Chromium turns into real JS touch events. The
pen/stylus stays on the wacom driver.
Removed the earlier dead-end attempts (xsetwacom MapToOutput / CTM reset,
Wacom Enable Touch Gesture, 99-wacom-touch.conf) which were all chasing the
wrong cause while the keyring grab masked any real testing. The upgrade path
removes the stale 99-wacom-touch.conf so it can't override libinput.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
The Wacom driver owns the Coordinate Transformation Matrix and silently
overrides any xinput set-prop changes. xsetwacom MapToOutput tells the
driver to recalculate the CTM for the primary connected output, which is
the correct API and persists across driver resets.
Dynamically detects the primary output (eDP1, HDMI1, DP1, etc.) so the
fix works on any machine without hardcoding a display name.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
The Wacom driver can initialise the CTM to all-zeros, which maps every
touch event to screen coordinate (0,0). The touchscreen appears completely
dead even though the hardware and kernel are working correctly.
Reset the CTM to the identity matrix for every touch/finger device at
startup, before launching Electron, so coordinates are always correct.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
xorg uses fnmatch (shell glob) for MatchProduct, where . is a literal
dot. Wacom.*Finger never matched "Wacom HID 48E3 Finger touch" because
there is no literal dot in that string. Wacom*Finger* matches correctly.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
Without /etc/X11/xorg.conf.d/99-wacom-touch.conf the Wacom driver initialises
the finger touch device in pointer emulation mode (generating RawButtonPress/
RawButtonRelease/RawMotion). Electron never sees TouchBegin/TouchEnd events so
touchstart/pointerdown(touch) never fire in the renderer.
Setting Option "Gesture" "on" and Option "Touch" "on" at the driver level means
the device initialises in XI2 touch mode on every X server start, regardless of
any post-init xinput set-prop calls.
Added to both fresh install (step 18) and upgrade function.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
Without XAUTHORITY set, xinput can fail with "Authorization required"
if the display manager doesn't propagate it through the session environment.
Hardcode the kiosk user's .Xauthority path to guarantee xinput works.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
Update all version strings (header, SCRIPT_VERSION, JS VERSION constant)
and rename ubuntu-based-kiosk-v1.0.2.sh → ubuntu-based-kiosk-v1.0.3.sh.
Update README with v1.0.3 change log and archive v1.0.2 as previous.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
The xinput loop now scans all input devices at startup, matches anything
with "touch" or "finger" in the name (excluding touchpads/trackpads), and
attempts to enable Wacom touch gestures on each match. Non-Wacom devices
silently ignore the set-prop call, so the loop is safe on any hardware.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
Add xinput call to start.sh so Wacom HID 48E3 touch gesture support is
initialized every time the kiosk starts, not just after lightdm restarts.
Also add start.sh to the upgrade extraction list so it is updated in place
instead of keeping the stale version from the original install.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
touchstart/touchend never fire on this device (confirmed by zero [TOUCH]
log entries). The activity tracker already uses pointerdown/pointerup and
works fine, proving PointerEvents reach the preload. Added pointer event
handlers that mirror the touch handlers for all gestures (2-finger swipe,
3-finger toggle, 1-finger arrow keys). A 500ms debounce on the IPC send
prevents double-firing on devices where both event types fire.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
capture:true and --touch-events=enabled were added to handle Authelia's
login page blocking touch events. Authelia now auto-logs in on startup
so the login page never shows. Reverting to the v1.0.0 approach (passive:true
only, no --touch-events flag) which had working two-finger swipe.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
With --ozone-platform=x11, Chromium defaults touch event detection to
'auto' and may not identify the hardware as a touchscreen, so touchstart/
touchend never fire in the renderer. --touch-events=enabled forces W3C
touch events on unconditionally, restoring two-finger swipe navigation.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
The upgrade was wiping node_modules then relying on npm to re-download the
~120MB Electron binary. npm returns exit 0 even when the download times out,
leaving the kiosk with no Electron binary and a blank screen on next boot.
node_modules only needs to be deleted on a fresh install or when explicitly
changing Electron version. For a JS-file-only upgrade, npm install without
a wipe is either a no-op (no changes) or applies dependency updates cleanly.
The install_electron_binary fallback remains as a safety net.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
Touch handlers used bubble phase (no capture:true), so any page script that
called stopPropagation() on touchstart/touchend — e.g. Authelia's login form
or scroll containers — silently blocked the preload's swipe detection.
Using capture:true fires the preload's listeners in the capture phase (before
any element-level handlers), so swipe works even on pages with their own
touch handling. Applied to both preloads (standard and auto-show keyboard).
Also adds missing [TOUCH] 2-finger HORIZONTAL console.log to the standard
preload so swipe events are visible in electron.log for debugging.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
The duplicate-block pitfall (YAML silently ignores duplicate keys, causing
a white screen) is now called out explicitly in both the script's printed
output and the README. Added a before/after example showing the correct
merged result with the kiosk one_factor rule above the two_factor wildcard.
Also explains why one_factor is required (TOTP/WebAuthn need interactive
second step, impossible via API).
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
Without a timeout, session.defaultSession.fetch() hangs for 1-2 minutes
on TCP timeout when Authelia is unreachable (wrong URL, server down,
firewall). Since createWindow() awaits autheliaAuthenticate(), the main
window is visible but no BrowserView is attached during that wait —
causing a persistent white screen with ibeam cursor.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
- Step 3 now says MERGE (not replace/append) with clear warning to keep existing config
- access_control: kiosk one_factor rule must go ABOVE any existing two_factor rule,
with explanation that Authelia applies rules top-down (first match wins)
- session block: keep existing values; only add the block if none exists yet
- Kiosk can only do one_factor — TOTP/WebAuthn via API is not possible
- Updated in both configure_authelia() printed output and README Authentication section
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
The upgrade_kiosk() path deleted node_modules then ran npm install but
never checked the binary or set chrome-sandbox permissions — so every
upgrade produced a blank screen.
Changes:
- Extract electron binary verification, fallback downloads, and
chrome-sandbox chmod 4755 into a shared install_electron_binary()
function called by both step 17/27 (fresh install) and step 5/6
(upgrade_kiosk) so neither path can silently skip the permission fix
- Add repair_electron() function: stops display, re-runs
install_electron_binary, restarts lightdm — no SSH needed
- Wire repair_electron as Advanced menu option 12 "Fix Blank Screen"
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
The ~120MB Electron binary download was silently failing because npm's
default 60s fetch timeout is too short on slower connections.
Changes to step 17/27:
- Set npm fetch-timeout to 600s and retries to 5 before running npm install
- If binary still missing after npm install, retry via install.js with
ELECTRON_FORCE_DOWNLOAD=true
- If still missing, fall back to direct wget download of the exact
versioned zip from GitHub releases (300s timeout, 3 tries, shows progress)
- Exit 1 with clear message if all three attempts fail
- Log chrome-sandbox permission step for visibility
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
/home/kiosk/ has 750 permissions so the install user can't read inside
it. The -f check and both jq reads were running as the current user and
failing silently, causing the false "config.json not found" error.
Changed:
[[ ! -f "$config_file" ]] → sudo test -f "$config_file"
jq -r ... "$config_file" → sudo -u kiosk jq -r ... "$config_file"
jq ... > "$tmp" → sudo -u kiosk jq ... > "$tmp" && sudo mv
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
npm install returns 0 even when Electron's postinstall binary download
fails, leaving node_modules/electron/dist/electron missing and causing
a blank screen with no useful error.
After npm install, explicitly check for the binary. If absent, retry
via ELECTRON_FORCE_DOWNLOAD=true node install.js. If still missing,
print a clear error and exit 1 instead of silently continuing to a
broken install.
https://claude.ai/code/session_01EyjEQLWbTXcZgbMDarf7NU
Claude