Move the build and deployment scripts into setup/tools alongside the existing
generators (proxmox-lxc-gen.sh, freeipa-image.sh, generate-modules.sh):
- setup/generate-answerfile.sh -> setup/tools/generate-answerfile.sh
- setup/archiso/build.sh -> setup/tools/build.sh
- setup/archiso/wds-deploy.sh -> setup/tools/wds-deploy.sh
- setup/archiso/write-usb.sh -> setup/tools/write-usb.sh
The archiso overlay/ (profile data: airootfs, profiledef.sh, packages.extra)
stays in setup/archiso/; build.sh now resolves it via an OVERLAY_DIR anchored
at the repo root, and DOTFILES_DIR is corrected for the new two-levels-deep
location. Updated every reference (generate-modules.sh, readme.md,
docs/md/{archiso,installation,index}.md).
generate-answerfile.sh: add linux-hardened to the kernel menu; verified its
emitted answerfile schema still matches every field the installers consume
(drive/kernel/keymap/hostname/username/password/luks_password/encrypt/
fido2_*/run_tui/components/desktop_environment/apps/shell_rc/colors).
Docs: add a dedicated ansipa Setup Guide (docs/md/ansipa-setup.md) covering
architecture, the full port list, the nginx portal + reverse-proxy paths, the
exact upstream reverse-proxy snippets (nginx/Caddy), NAT/push-mode behaviour,
and the Arch autofs/AUR caveat; link it from the index and cross-reference the
CheckMK/nginx ports from freeipa-ansible.md.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUhrcFU8J1Hnf7vNqNxZNi
Statically audited every optional-module and DE script's package tokens
against the official repos, the AUR, and Flathub. Ten modules named a package
that does not exist at the source they used, so each would abort with "target
not found" (pacman), an AUR miss, or a Flathub 404:
Installed via pacman but AUR-only (switched to the yay AUR helper):
- tetris: bastet
- doom: chocolate-doom, freedoom
- openarena:openarena
- openshot: openshot
- freeipa-client (module): oddjob — now optional via yay with a
pam_mkhomedir fallback (matches freeipa-enroll.sh)
Wrong AUR/repo package names:
- cecilia: cecilia -> cecilia-git
- productivity: python-jrnl -> jrnl (in the official repos)
- localtunnel: AUR fallback localtunnel -> nodejs-localtunnel
Wrong/absent Flatpak IDs:
- zen-browser: io.github.zen_browser.zen -> app.zen_browser.zen
- stuntrally: io.github.stuntrally.StuntRally3 is not on Flathub at all;
switch to the AUR stuntrally-bin (the actual Stunt Rally 3)
Verified all corrected names resolve (repo / AUR RPC / Flathub API). Modules
whose tokens are in the official repos and merely installed via yay (nyxt,
toot, llama-cpp — yay handles both sources) were left as-is.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUhrcFU8J1Hnf7vNqNxZNi
Add an nginx service that fronts all three web UIs on a single hostname/port
and serves a portal landing page, designed to sit behind the operator's own
TLS-terminating reverse proxy.
- Each backend is proxied at the path it already serves natively — FreeIPA
/ipa, CheckMK /cmk, Keycloak under a configured /auth relative path — so no
fragile path rewriting is needed and one upstream proxy line exposes
everything. FreeIPA's strict Host/Referer anti-CSRF checks are satisfied by
pinning Host to the IPA hostname and rewriting Referer.
- Honors incoming X-Forwarded-Proto/For so it works behind a second reverse
proxy (double-proxy verified: portal, FreeIPA, CheckMK and Keycloak all
reachable through two layers, and a CheckMK login POST completes to an
authenticated session through the gateway).
- Keycloak: KC_HTTP_RELATIVE_PATH=/auth + KC_PROXY_HEADERS=xforwarded so it
builds correct URLs behind the gateway; deliberately not forwarding
X-Forwarded-Port (this gateway is :80 internally; the public port is the
upstream proxy's) to avoid https://host:80 redirects.
- .env.example documents ANSIPA_HTTP_PORT and gives the exact upstream
reverse-proxy snippet (nginx location block + Caddy one-liner).
Portal is a static cyberqueer-themed page linking to /ipa/ui/, /cmk/, /auth/.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUhrcFU8J1Hnf7vNqNxZNi
Found while running the policy matrix and NAT/WAN scenario in VMs:
- ansipa-enforce-policies.sh and the dev_security-scan generated script called
`hostname -f`, but the hostname binary (inetutils) is not installed on a
minimal Arch system, so the enforcer aborted with "hostname: command not
found" on every run. Add an ansipa_fqdn() helper (hostnamectl / getent /
/etc/hostname / kernel-hostname fallback chain) and use it; inline the same
fallback in the standalone scan script.
- Fix two stacked bugs that meant CheckMK push mode — the only monitoring
transport that works when a client is behind NAT — never functioned:
* ansipa-checkmk-setup.sh referenced IPA_ADMIN_PASS in the DONE_FLAG
early-exit path, but that variable was assigned only afterwards, so every
rerun kinit'd with an empty password, failed silently, and skipped the
push-mode request writer. Resolve config before the early exit.
* _write_push_requests parsed `ipa hostgroup-show dev_mon_base` members with
awk that did `next` on the "Member hosts:" line — but ipa prints members on
that same line, so the only line with hostnames was discarded and no push
request was ever written. Rewrite the awk to capture the label line's value
plus wrapped continuation lines. Verified end-to-end: requests are written
and CheckMK flips the host to cmk_agent_connection=push-agent.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUhrcFU8J1Hnf7vNqNxZNi
core.sh enables greetd (tuigreet) on every install, but the ly-based Hyprland
DEs only disabled getty@tty1 and enabled ly@tty1 — leaving both display
managers enabled and racing for tty1, so the stale greetd/tuigreet greeter
could win and show the old text login instead of ly.
Add a disable_service helper (mirrors enable_service; best-effort, non-fatal)
and call disable_service greetd.service in hyprlua.sh and hyprland.sh right
before enabling ly@tty1, so ly owns the login screen.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUhrcFU8J1Hnf7vNqNxZNi
Found while bringing up the FreeIPA+CheckMK+Keycloak stack in a VM:
- Pin volume and network names (name:) so compose and the cgroup-v2 run.sh
wrapper share the same objects. Compose otherwise prefixes the project name
(image_cmk-creds, image_ipa-net), so the run.sh-started freeipa container
mounted a different cmk-creds volume than the compose-started checkmk wrote
to — the automation secret never reached IPA and the dev_mon_* integration
retried forever. run.sh also now labels the network it creates so compose
adopts it instead of erroring "exists but not created by compose".
- run.sh: mount cmk-creds into freeipa and include checkmk in the "all" target;
both were missing, so the CheckMK path was never wired up via run.sh.
- Fix both healthchecks: Keycloak has no curl/wget and serves /health on the
management port 9000 only when KC_HEALTH_ENABLED=true, so the old curl probe
left it permanently unhealthy — use a bash /dev/tcp probe against 9000.
CheckMK's /api/1.0/version needs auth (401), so probe the login page instead.
- Advertise a client-routable CheckMK URL (CMK_ADVERTISED_URL) in the
dev_mon_base description instead of the container-internal 172.30.0.12, which
enrolled clients cannot reach; falls back to the internal URL when unset.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUhrcFU8J1Hnf7vNqNxZNi
Found while testing client enrollment across distros in VMs:
- Restore the executable bit on freeipa-enroll.sh (lost on a prior edit), which
made freeipa-client.sh fail with a misleading "not found" for every path.
- Arch: freeipa-client is AUR-only, not the repo package "freeipa" — pacman -S
freeipa died with "target not found". Install sssd/krb5 from repos and
freeipa-client via an aur_install helper that drops to the sudo user.
- Bundle autofs-pkgbuild/: freeipa-client (AUR) hard-depends on autofs (AUR),
whose PKGBUILD pins autofs-5.1.9.tar.xz — pruned from kernel.org at the 5.2.0
release, so it 404s and blocks all Arch enrollment. Build autofs ourselves
from upstream git pinned to a 5.1.9-series commit that already carries the
cyrus-sasl function-pointer fix GCC 14+ needs, and install it before the
freeipa-client AUR build so the broken AUR autofs is never touched. Verified
the bundled PKGBUILD builds and installs cleanly on Arch.
- Fedora vs RHEL/Rocky/Alma use different client package names: keep
freeipa-client on Fedora, use ipa-client on the RHEL family (freeipa-client
does not exist there). Split the case arms accordingly.
- Guard both freeipa-client.sh entry points against unattended/no-TTY runs: the
module TUI would hang on a dialog menu, and the wrapper would block on an
interactive password read when the shipped answerfile has a blank password.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUhrcFU8J1Hnf7vNqNxZNi
Found while testing both installers end-to-end in VMs:
- Wait for archiso's pacman-init.service instead of racing it. The installer
autostarts at login while the live keyring is still being populated
asynchronously, so every package failed with "signature is unknown trust" /
"invalid or corrupted package (PGP signature)". Do not run pacman-key --init
ourselves (corrupts the half-built keyring); wait, then refresh
archlinux-keyring so an ISO built weeks ago still trusts current keys.
- archbaseos-guided-install.sh: honor the answerfile's password/luks_password
fields; previously it always prompted interactively, so a documented
"zero-interaction" netboot/WDS install hung forever with no console.
- Install GRUB to the UEFI removable-media fallback (/EFI/BOOT/BOOTX64.EFI) in
addition to the NVRAM entry. Without it, an installed disk with empty/cleared
NVRAM (fresh board, CMOS reset, or a cloned/redeployed image — the whole point
of this installer) shows "No bootable device". Verified: the disk now boots to
the login manager with fresh NVRAM.
- build.sh: sudo rm -rf the work dir. mkarchiso runs as root and leaves
root-owned files if interrupted, so the next plain rm -rf aborted the script
under set -e before mkarchiso could run.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUhrcFU8J1Hnf7vNqNxZNi
- Use samba/openssh package names on Arch instead of Debian-style samba-client/openssh-server
- Skip AUR-only freeipa install when ipa-client-install already present; skip oddjob if not in official repos
- Guard seed_policystore delegate_to when ipa_server_host is undefined
- Make hostname detection and OS version parsing more robust
- Fix inverted DNS_UPDATE/CONFIGURE_SUDO enrollment flags
The 1550-line ansipa-enforce-policies.sh is replaced by a slim orchestrator
that sources policy files from policies.d/ — one file per policy type — via
a shared lib/ansipa-policy.sh that consolidates all IPA group discovery into
two queries (one host-show, one group-find) instead of five or more.
New files:
- ansible/lib/ansipa-policy.sh shared constants, log/warn, _ansipa_discover(),
_smb_parse_cred(), _ansipa_sync_policystore()
- ansible/policies.d/dev_daemon.sh
- ansible/policies.d/dev_timeshift-backup.sh
- ansible/policies.d/dev_security-scan.sh
- ansible/policies.d/dev_no-local-users.sh
- ansible/policies.d/dev_local-sudo.sh
- ansible/policies.d/dev_ssh.sh
- ansible/policies.d/dev_mon.sh includes usr_mon_logins (same CMK dep)
- ansible/policies.d/usr_admin.sh
- ansible/policies.d/usr_block-binary.sh
- ansible/policies.d/usr_prt.sh
- ansible/policies.d/usr_scan-notify.sh
- ansible/policies.d/usr_smb.sh
- ansible/policies.d/usr_policystore.sh NEW: auto-mount ~/policystore for
usr_smb_adm_policystore members
The policies.d/ folder on the FreeIPA container is exposed as an SMB share
(ansipa-policystore) so admins can live-edit policy files without Ansible.
The enforcer syncs from the share before each run; cached local files are
used as fallback when SMB is unreachable.
- ansipa-smb-setup.sh: adds ansipa-policystore share + smb-policystore system
user + usr_smb_adm_policystore IPA group with auto-generated credential
- deploy-ansipa-policies.yml: deploys lib/ and policies.d/ to enrolled clients;
adds optional seed_policystore task for bootstrap
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- ansipa-enforce-policies.sh header: add missing dev_ssh_*, usr_smb_r_*,
usr_smb_rw_*, dev_mon_* (base/malware/timeshift/power), usr_mon_logins
with concise descriptions matching actual behaviour. Notes section
updated with CheckMK setup prerequisite.
- ansipa_packages local check: replace broken rpm||dpkg fallback (wc -l
always exits 0 so dpkg/pacman branches were never reached) with an
if/elif/elif that probes for rpm, dpkg, or pacman in order. Includes
pacman.log-based recent-packages list for Arch hosts.
- deploy-ansipa-policies.yml prerequisites: add CheckMK and SMB mount notes.
- .env.example: add CheckMK section (CMK_ADMIN_PASSWORD, CMK_SITE_ID).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds full CheckMK CE 2.3 monitoring integration to the ansipa policy
enforcement system. Monitoring is group-driven — adding a host to an
IPA hostgroup installs the relevant agent checks automatically.
New monitoring groups:
dev_mon_base — install check-mk-agent, register host in CMK;
packages check + built-in CPU/RAM/disk/uptime
dev_mon_malware — ClamAV scan results from dev_security-scan log
dev_mon_timeshift — Timeshift snapshot age (WARN >5d, CRIT >10d)
dev_mon_power — CPU package TDP via Intel RAPL or lm-sensors
usr_mon_logins — SSH login attempts (success/failed/invalid-user)
Key implementation details:
- ansipa-checkmk-setup.sh: one-time folder/group creation on FreeIPA
container; reads automation secret from shared cmk-creds volume;
stores credentials in IPA dev_mon_base description for clients
- ansipa-checkmk.service: starts after IPA first-boot, retries until
both IPA and CheckMK are ready
- docker-compose.yml: CheckMK container adds port 8000 (agent receiver
for push mode); command override writes automation.secret to shared
volume, patches cookie_auth.conf for API access, and polls
/cmk-creds/push-mode-*.req to configure push-agent hosts in hosts.mk
- ansipa-enforce-policies.sh: local check scripts for all 5 groups;
cmk-agent-ctl registration + auto push timer when push mode is active
- CheckMK REST API path corrected to /cmk/check_mk/api/1.0/ (not
/cmk/api/1.0/ which returns 404 in 2.3)
- .gitignore added to prevent accidental commit of .env test creds
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DHops6PU4c2Mv5UyhUj8Ms
usr_smb_r_<name> / usr_smb_rw_<name> (user groups):
- Server creates /data/smb-shares/<name>/ with smb-<name> system user,
setgid directory, and Samba share stanzas [usr-<name>-r/rw] on startup
- Auto-generates credentials and stores them in IPA group descriptions
(cifs://<host>:<samba-user>:<password>) so no out-of-band secret distribution
- Client mounts //ipa/<share> at ~/‹name› for logged-in members;
rw membership wins over r when user is in both groups
- Revert: unmounts and rmdir the mount point on group leave
dev_ssh_<userid> (host group):
- Reads IPA user's SSH public keys (ipasshpubkey / --all output)
- Writes them into /home/<userid>/.ssh/authorized_keys in an
ansipa-managed section (preserves manually-added keys)
- Home dir is created if it doesn't exist (user may not have logged in yet)
- Revert: removes only the ansipa-managed section on group leave
- Enables same-key SSH access from any registered device to enrolled hosts
Also: add cifs-utils + openssh-server to Ansible package list
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DHops6PU4c2Mv5UyhUj8Ms
Three bugs fixed, all confirmed end-to-end in a live FreeIPA + client VM:
1. Wrong `who` column: loginctl is SESSION/UID/USER (col 3), but `who` is
USER/TTY/DATE (col 1) — the fallback used col 3, delivering to the wrong
user. Split into _parse_session_user() with per-source column selection.
2. Delivered to all users: alerts were sent to every active non-root session
instead of only members of usr_scan-notify. Added `id -nG` group check
inside _parse_session_user().
3. smbclient ls path breakage: `smbclient -c "ls alerts\host\"` treats the
argument as a glob pattern, returning NT_STATUS_NO_SUCH_FILE. Changed to
`cd alerts\host; ls` which correctly lists directory contents.
Bonus: track ACKNOWLEDGED set so alerts deleted-from-server in the current
run are not re-downloaded (avoiding a spurious WARN on the same run).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DHops6PU4c2Mv5UyhUj8Ms
Fixes found during a complete archiso → FreeIPA DC → client enrollment →
ansipa policy deployment test cycle (Rocky 9 client VM, FreeIPA in Docker):
arch-autoinstall.sh:
- Add openssh to pacstrap — was missing, breaking headless/automated installs
- Enable sshd at boot and set PermitRootLogin yes for post-install access
ansipa-install-packages.sh:
- Fix broken uninstall logic: three always-true conditions caused every package
removal to be silently skipped; replace with a state-file-based approach that
tracks packages installed by ansipa and removes only those when the IPA group
disappears
ansipa-enforce-policies.sh:
- Add usr_admin policy: creates /etc/sudoers.d/ansipa-usr-admin granting full
sudo to the FreeIPA usr_admin user group on every enrolled host; reverted
when the IPA group is deleted
- Add usr_prt_<printer> policy: auto-adds CUPS printers for FreeIPA user group
members at login; printer URI stored in IPA group description; per-user ACL;
reverted when group is deleted or membership ends
deploy-ansipa-policies.yml:
- Document usr_admin and usr_prt_* policies in header comment
- Install cups package on clients (required for printer policy)
FreeIPA container (image/):
- Add docker-env.service: extracts IPA_*/SMB_*/LUKS_*/KEYCLOAK_* env vars from
/proc/1/environ into /etc/container.env on container start; services read from
there rather than relying on PassEnvironment (which is lost on container restart)
- Add run.sh: wrapper that starts FreeIPA with --cgroupns host, required on cgroup
v2 hosts (WSL2, recent Linux) because Docker Compose schema does not expose
cgroupns_mode; also wires LUKS_KEY_UPLOAD_PASSWORD which the SMB service needs
- Dockerfile: copy + enable docker-env.service; add glibc-langpack-en for locale
- ansipa-smb.service: use EnvironmentFile=/etc/container.env (via docker-env.service)
instead of PassEnvironment; add docker-env.service dependency
- ipa-first-boot.service: add docker-env.service dependency + EnvironmentFile
- ipa-first-boot.sh: add --skip-mem-check (container has limited cgroup memory
visibility); remove --no-reverse from non-DNS path (was invalid without DNS)
- docker-compose.yml: add prominent cgroupns note explaining when to use run.sh
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DHops6PU4c2Mv5UyhUj8Ms
Diagnosed live in a VM: the bar rendered twice (two stacked rows). `hyprctl
layers` showed two `gtk-layer-shell` bar surfaces owned by two different eww
pids — an `eww daemon` and an `eww open bar`. The monitor count was correctly 1,
so the per-monitor loop was not at fault.
Root cause: ewwstart.sh started `eww daemon` and then immediately ran `eww open`
in a loop. The first `eww open` frequently ran before the freshly-started daemon
was ready, so it spawned its OWN second daemon and drew a second bar.
Fix (all three DE variants — hyprlua, hyprland, niri):
- Drop the separate `eww daemon`; rely on `eww open` to auto-start the daemon and
block until the window is mapped, so the first call establishes a single daemon
that the rest reuse (no race, no second daemon).
- killall eww + short settle before opening, so a reload starts clean.
- hyprlua/hyprland: count monitors via `hyprctl monitors -j | jq length` instead
of the fragile `hyprctl monitors | grep ID | wc -l`.
- export GTK_THEME (the bare assignment never took effect).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R5kHioUMK3mtf2eiLEozCM
Same Vicinae theming bug as hyprlua, now fixed across the remaining DE modules:
- Vicinae loads named custom themes only from ~/.local/share/vicinae/themes/*.toml,
not ~/.config/vicinae/. Both modules now deploy cyberqueer.toml there so
settings.json's theme.dark.name="cyberqueer" resolves instead of falling back
to the stock dark theme.
- hyprland: the source had settings.json double-nested at vicinae/vicinae/, so it
deployed to ~/.config/vicinae/vicinae/settings.json and was never read. Flattened
the source to vicinae/{settings.json,cyberqueer.toml} (matching hyprlua), so it
now lands at ~/.config/vicinae/settings.json.
- niri/vicinae is a symlink to hyprlua/vicinae and niri.sh uses `cp -rL`, so its
settings.json was already correct; only the theme-deploy line was needed.
Verified by simulating both deploys into a temp HOME: final layout matches the
known-good hyprlua result.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R5kHioUMK3mtf2eiLEozCM
The config-copy step put cyberqueer.toml into ~/.config/vicinae/, but Vicinae
loads NAMED custom themes only from ~/.local/share/vicinae/themes/*.toml (per
the official docs). So settings.json's theme.dark.name="cyberqueer" never
resolved and the launcher silently fell back to the stock dark theme.
Copy the theme into ~/.local/share/vicinae/themes/ so the named theme resolves
and the cyberqueer palette is applied.
Note: the hyprland and niri DE modules have separate, pre-existing Vicinae
layout issues (hyprland double-nests settings.json; niri bundles no vicinae
config) — left for a dedicated follow-up.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R5kHioUMK3mtf2eiLEozCM
core-packages installs ffmpeg, which depends on the virtual `jack`. With only
`pipewire` (not pipewire-jack) requested, pacman --noconfirm picks the default
provider `jack2`. The DE modules then install `pipewire-jack`, which conflicts
with the already-installed jack2 — pacman can't auto-resolve it and aborts the
whole DE module ("pipewire-jack and jack2 are in conflict").
Explicitly list pipewire-jack/pipewire-pulse/pipewire-alsa in core-packages so
PipeWire is the chosen provider from the start; jack2 is never pulled and the DE
package install no longer conflicts.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R5kHioUMK3mtf2eiLEozCM
Two more module aborts surfaced by the live VM test (both pre-existing, only
unmasked once the wprs bug was fixed and the transaction got further):
1. core-packages.sh listed BOTH `rust` and `rustup`. They conflict (both provide
cargo/rustc); pacman --noconfirm cannot auto-resolve it and aborts the whole
core-packages transaction ("unresolvable package conflicts"), so wget and the
other core tools never installed. Drop `rust` — the setup uses rustup
exclusively (`rustup default stable`).
2. core.sh copied the greetd config into /etc/greetd before that directory
existed (greetd is pulled in later by the DE module), so the cp failed and
`set -e` aborted Core Services before fail2ban/udisks2 were enabled. mkdir -p
/etc/greetd first.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R5kHioUMK3mtf2eiLEozCM
Live VM test revealed the HyprLua DE install aborted at the cursor-download step
("wget: command not found", exit 127), leaving ~/.config/hypr undeployed. Two
root causes:
1. core-packages.sh listed `wprs` in the pacman array, but wprs is AUR-only
(wprs-git). pacman aborts the ENTIRE transaction on an unknown target, so the
whole core-packages install failed with "target not found: wprs" — and NONE
of the core packages installed, including wget. Removed wprs (it has its own
optional wprs.sh AUR module) and documented why it must never be re-added.
2. hyprlua/hyprland/niri downloaded the cursor theme and wallpaper with wget,
which is not part of the base system. Switched them to curl (always present),
so the DE install no longer depends on an earlier module having installed wget.
Without these, HyprLua silently half-installed (packages yes, config no) and the
booted system had no Hyprland configuration.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R5kHioUMK3mtf2eiLEozCM
A bare `chsh -s /usr/bin/zsh` authenticates the calling user through PAM and
prompts "Password:", stalling an unattended install right after oh-my-zsh.
Run it via `sudo chsh ... "$(whoami)"` so it uses the passwordless setup sudo
rule and completes silently (also more reliable: it targets the account
explicitly rather than relying on the ambient user).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R5kHioUMK3mtf2eiLEozCM
oh-my-zsh's install script, even with RUNZSH=no/CHSH=no, still prompts
"Do you want to overwrite it with the Oh My Zsh template? [Y/n]" in interactive
mode — which hangs an unattended/answerfile install at the shell-setup module.
Pass --unattended to skip all prompts, and KEEP_ZSHRC=yes so it preserves the
dotfiles ~/.zshrc symlink (created earlier in this module) instead of replacing
it with the oh-my-zsh template.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R5kHioUMK3mtf2eiLEozCM
Audited every module for prompts that would hang an answerfile/unattended run.
The EWW form-factor question and several interactive-config modules were still
unguarded (only hyprlua had been fixed):
- hyprland.sh, niri.sh: skip the EWW form-factor `read` when MARCHY_UNATTENDED=1
or stdin is not a TTY, defaulting to the desktop/no-battery bar (matches hyprlua).
- mail-notmuch.sh, caldav-sync.sh: install the tools, then exit cleanly in
unattended mode instead of blocking on the account/server credential prompts —
the user configures the account after first boot.
- freeipa-server.sh: bail out early in unattended mode; FreeIPA server provisioning
is interactive and must run on a booted system (ipa-server-install needs running
services), so it can never run during the install.
freeipa-client.sh is left as-is: it has a genuine --unattended enrolment path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R5kHioUMK3mtf2eiLEozCM
The temporary setup drop-in granted `NOPASSWD: ALL`, which covers `sudo <cmd>`
but NOT `sudo -v`. Installers run by the TUI (starship, rustup, …) call `sudo -v`
to pre-authorise, and that check still demands a password whenever the user has
any password-required sudoers entry — which they do, via the wheel rule in
10-wheel. The result was a hidden `[sudo] password for <user>:` prompt that
stalled the otherwise-unattended module install.
Add `Defaults:<user> !authenticate` to the 99-setup-nopasswd drop-in (in both
the auto and guided installers) so the auth check is skipped entirely for the
setup user; `sudo -v` and `sudo <cmd>` are now both passwordless during setup.
Verified live in a VM: `sudo -nv` for the user went from "a password is required"
to rc=0 after adding the line.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R5kHioUMK3mtf2eiLEozCM
Previously the user password (and the LUKS passphrase for encrypted installs)
were always prompted interactively, so an answerfile install could never be
fully hands-free. Add optional "password" and "luks_password" answerfile fields:
- arch-autoinstall.sh: read both via af_get; when present use them (chpasswd /
cryptsetup --key-file=- with --batch-mode and stdin-piped luksAddKey auth),
otherwise fall back to the interactive prompt. Empty/null/absent => prompt.
- generate-answerfile.sh: replace the "passwords are never stored" notice with
an optional confirmed-entry password prompt (and a LUKS one when encryption is
enabled); emit both as JSON (null when declined).
Secrets stored this way are plain text in the file (and world-readable once
embedded in an ISO) — documented in the header; decline to keep prompting.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R5kHioUMK3mtf2eiLEozCM
An answerfile/PXE install runs the TUI modules with no operator present, so any
interactive prompt or confirmation hangs the deployment forever.
- tui-install.sh: run each module in the FOREGROUND with a background `tail -f`
mirroring its log, instead of backgrounding the module. Backgrounding made
interactive module reads (e.g. the EWW form-factor question) trigger SIGTTIN
and stop; the file-based log still avoids the `| tee` pipe-inherit hang that
paused installs after a module spawned a daemon. Export MARCHY_UNATTENDED=1
in answerfile mode so modules can detect it.
- hyprlua.sh: skip the EWW form-factor `read` when MARCHY_UNATTENDED=1 or stdin
is not a TTY, defaulting to the desktop/no-battery bar; add --noconfirm to the
tablet-branch yay call.
- python.sh: add --noconfirm --needed to its `pacman -Syu` so it no longer waits
on the "Proceed?" prompt during unattended installs.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R5kHioUMK3mtf2eiLEozCM
When the TUI modules run inside the archiso installer chroot, the new
system's systemd is not the running init and there is no user session bus.
Operations like `systemctl start`, `enable --now`, `systemctl --user`, and
`gsettings` fail there and, under `set -e`, abort the whole module.
- logging.sh: add in_chroot/have_user_bus/enable_service/start_service
helpers. enable_service warns instead of aborting; start_service skips in
a chroot (unit starts on first boot via its enable symlink).
- core.sh, hyprland.sh, hyprlua.sh, niri.sh: route enables through
enable_service, starts through start_service, and guard gsettings behind
have_user_bus. Fixes the cronie-enable failure and the ly/DM setup abort.
- app modules (tlp, timeshift, open-webui, mysql, qemu, ollama, cockpit,
docker, ssh-server): convert `enable --now`/plain enables to
enable_service + start_service so they no longer abort during chroot install.
- tui-install.sh: run modules with output to a file plus a pid-bound tail,
waiting on the module PID, so a daemon child inheriting the pipe can no
longer hang the installer after a module (e.g. flatpak) finishes.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R5kHioUMK3mtf2eiLEozCM
The base installers granted wheel sudo with `echo '%wheel ... ALL' >> /etc/sudoers`.
Because the stock sudoers ends with `@includedir /etc/sudoers.d`, that appended
rule is parsed AFTER the drop-ins, and since sudo applies the last matching rule,
it overrode the temporary 99-setup-nopasswd NOPASSWD rule — so the user had to
re-enter their password on every pacman/yay/flatpak call while the TUI installed
modules.
Grant wheel via /etc/sudoers.d/10-wheel instead, which sorts before
99-setup-nopasswd so NOPASSWD wins during the TUI run and password auth resumes
once the temp file is removed. Also guard that @includedir is present (so the
drop-ins are always read) and set both drop-ins to the canonical 0440 mode.
Applied to both archbaseos-guided-install.sh and arch-autoinstall.sh.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Flip their default field to "on" in modules.conf and regenerate the checklist
blocks in tui-install.sh and generate-answerfile.sh so they come pre-selected
in both the TUI installer and the answerfile generator.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Without oflag=direct, dd copies the ISO into the page cache at high speed, the
status=progress counter races to ~100%, then appears to hang at the end while
the slow USB drains and conv=fsync flushes — looking frozen when it is still
writing. write-usb.sh now writes with oflag=direct (honest device-level
progress) and falls back to a cached write if the bridge rejects O_DIRECT;
build.sh's dd hint and explanation are updated to match.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
"failed to load ldlinux.c32" when booting from USB is caused by writing the
isohybrid ISO at file level (drag-and-drop, Rufus "ISO" mode, UNetbootin), which
destroys the MBR/isolinux layout. The ISO build itself is correct — archiso's
bios.syslinux mode installs isolinux.bin + isohdpfx.bin + the c32 modules and
applies the isohybrid MBR, and the profile's bootmodes match upstream releng.
Add archiso/write-usb.sh: a safe block-level (dd) writer that lists removable
disks, refuses partitions and the system/root disk, requires an all-caps YES,
unmounts the target, then writes with conv=fsync. build.sh's completion output
now points at it and warns that file-level copies cause exactly this error.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
shellcheck passes (default severity) on the base installer and the whole
answerfile chain: generate-answerfile.sh → build.sh → .automated_script.sh →
launch.sh → arch-autoinstall.sh → tui-install.sh. Verified the generator's
JSON schema (drive, kernel, keymap, hostname, username, encrypt, fido2_root,
fido2_user, run_tui, components, desktop_environment, apps, shell_rc, colors)
parses correctly through both the auto base installer (af_get/af_bool) and the
TUI installer (load_answerfile), for full and minimal/empty answerfiles.
Fixed the three legitimate findings surfaced along the way:
- generate-answerfile.sh: drop unused AVAIL_DRIVES (SC2034); the drive list is
rendered inline in the dialog prompt.
- tui-install.sh: drop unused C_RULE (SC2034); write log truncation as
': > "$LOG"' so the redirection has an explicit command (SC2188).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
In answerfile mode the hostname now always gets the machine's MAC appended as
"<name>-<mac>" with the MAC stripped of all separators (colons/dots/dashes), so
fleet deployments from one answerfile stay unique. The default "arch" name now
also receives the suffix, and the dash is only added when a MAC is actually
found (no trailing "-" on NIC-less machines).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Flatpaks now install globally again: ensure_flatpak adds the Flathub remote at
--system and all 19 app modules use `sudo flatpak install --system`. Running
via sudo (root) performs the system op directly, avoiding the SystemHelper/
polkit D-Bus path that caused "The name is not activatable" for non-root users.
- tui-install.sh no longer prompts for or sets the hostname — the base installer
already configures it. Removed the Hostname section, the MAC-suffix helper, the
AF_HOSTNAME field and the summary line.
- archbaseos-guided-install.sh now gathers ALL input up front, including
passwords. New ask_password() prompts in clear text (by request) and requires a
confirmation entry, looping until the two match — so each password is typed
exactly twice and never again. The LUKS passphrase is captured once and fed to
luksFormat/open/luksAddKey (--key-file=-) and cryptenroll ($PASSWORD), instead
of cryptsetup prompting repeatedly. After all input, a single all-caps "type
YES" gate replaces the old per-step confirmations (answerfile mode keeps its
5-second abort window). The run-TUI choice is also asked up front.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Diagnosed from a full guided-install log plus a Hyprland startup log. Three
distinct remaining failures:
1. ~/.config left root-owned. The FIDO/Yubico setup runs `mkdir -p
~/.config/Yubico` as root (creating ~/.config itself), then chowned only
Yubico/. ~/.config stayed root-owned, so every later user step failed with
EACCES: shell-setup symlinks (starship.toml), the mail/caldav systemd --user
timers, and Hyprland creating ~/.config/hypr at startup. Chown the whole
~/.config in both Yubico spots, and defensively reclaim it in shell-setup.
2. python/wprs/plymouth/zfs sourced ../lib/logging.sh, but apps/ modules need
../../lib — so they aborted with "No such file or directory". Corrected.
3. Flatpak app modules ran `flatpak install -y` at system scope, which needs the
Flatpak SystemHelper D-Bus service + polkit (unavailable in a chroot/TTY
install) — the "The name is not activatable" failures (wireshark, xournal,
rnote, firefox-browser, …). Switch ensure_flatpak and all 19 main-flow
installs to --user scope, matching apply_flatpak_theme's --user overrides.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The core, hyprlua and hyprland package installs used backslash-continued
`pacman` commands with inline `#` comments. Bash treats the first inline
comment as the end of the command, so pacman ran with only the packages
before it and every later name (7zip, cronie, nwg-dock-hyprland, …) was
executed as a shell command — failing under `set -e`. Move all three lists
into arrays, where per-item comments are valid, and install with `--`.
Also:
- himalaya: install the official `himalaya` package (AUR `himalaya-bin` is gone).
- mail-notmuch / caldav-sync: make the systemd *user* timer setup and the
initial sync best-effort. A bare TTY/chroot install has no user session bus
(and ~/.config may not be writable yet), so `systemctl --user` and the mkdir
could abort the module; warn and continue instead.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The applications menu (~80 entries) overflowed the terminal. Render
the checklist one terminal-height page at a time, clearing and homing
the cursor each round for a stable scrollable view, with n/p paging.
Item numbering stays global so any entry can be toggled from any page.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Drop the dialog dependency entirely so the installer runs on a bare
console with only bash + coreutils. Reimplement the needed widgets
(msgbox, yesno, input, menu, checklist, form) as ui_* helpers using
read, preserving the cyberqueer magenta/cyan palette via ANSI codes
and the stdout/stderr fd convention so existing capture sites work
unchanged. Update generate-modules.sh to emit the ui_checklist form.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
opendeck is not in the official repos, so install the native AUR package
with yay (matching the repo's other AUR modules) and drop the Flatpak
path. Wire `opendeck` into Hyprland autostart instead of
`flatpak run com.mairtech.OpenDeck`, and update the idempotency guard.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Target ~/.config/hypr/usr/monitors.lua (the config Hyprland actually
loads) so `hyprctl reload` applies layouts immediately; the Dotfiles
repo copy stays the deploy source and is no longer written.
Add overlap geometry helpers and integrate them into the apply flow:
- block moves that would drive a monitor into a neighbor (TUI coords)
- snap positions to the MOVE_STEP_FINE grid to avoid frozen digits
- auto-resolve snap-induced collisions by re-reading the live layout
and nudging the moved monitor clear, up to MAX_RESOLVE_ITERS
- warn on residual overlap after apply and after save/reload
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace the machine-specific DP-1/DP-3/HDMI-A-1 layout with a wildcard
catch-all rule (preferred mode, auto position, auto scale) that brings up
any connected output. Per-machine layout stays the job of monitor-manager,
which overwrites this file.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
#2 — Word-boundary match for all module patterns
Generator now pads SELECTED_APPS with spaces and uses *" id "* in
counters, summary, and dispatch, matching the conflict fix from #1.
plymouth-custom no longer false-triggers any plymouth check.
#3 — Guided installer now runs tui-install.sh
archbaseos-guided-install.sh was calling simple-install.sh; both
paths now use the full TUI (sentinel-managed, modules.conf-driven).
#4 — EFI/boot partition size unified at 10 GiB
arch-autoinstall.sh was 15 GiB, archbaseos-guided-install.sh was
5 GiB. Both now use 10 GiB.
#5 — Interactive retry for dotfiles clone (guided installer)
Clone moved outside the chroot heredoc so read() reaches the terminal.
Loops until success or the user skips; AF_MODE warns and continues.
#6 — PAM target unified on system-local-login
archbaseos-guided-install.sh was writing to system-auth (affects
sudo). Both installers now target system-local-login only.
#7 — Redundant second clone removed from autoinstaller
arch-autoinstall.sh had a second git clone inside the chroot as a
fallback that collided with the skel copy and printed a spurious
warning. Removed; skel-only approach matches the guided installer
(last updated). Also removed the individual .zshrc/.bashrc/.vimrc
cp block; aligned to the guided installer's cleaner skel structure.
#8 — Docs: remove stale plymouth core-module section
docs/md/modules.md still described plymouth under Core Modules.
Section removed; plymouth appears in Optional Applications (system
category) via the generated sentinel.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two bugs in the module-conflicts system:
1. warn() was called by the generated conflict block but never defined —
any conflict would crash with 'warn: command not found'.
Added warn() to helpers: dialog msgbox in interactive mode, logged
printf in answerfile mode.
2. Conflict patterns used substring globs (*"id"*) which caused
plymouth-custom to match the plymouth check — selecting only
plymouth-custom would trigger the conflict block, call the missing
warn(), and then remove plymouth-custom from SELECTED_APPS, leaving
no boot splash running at all.
Fixed by padding SELECTED_APPS with spaces and using *" id "* word-
boundary patterns in both the condition and the removal substitution.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
