Local LLM server on the 6800 XT (ollama-rocm, gfx1030 needs no
HSA_OVERRIDE_GFX_VERSION) fronted by a LibreChat web UI, talking to it
over the OpenAI-compatible /v1 route. Also wires up LibreChat's
persistent-memory feature, which needed its own agent+model plus a
custom extraction prompt: the default 3b model couldn't reliably tell
the user's stated facts apart from its own boilerplate, and even a
tuned prompt didn't fix that — so memory extraction now reuses
gemma4:12b, the same model as the daily-driver chat endpoint.
flake.lock bump pulls in the ollama and librechat NixOS modules.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
After the rack move one of the RAID0 disks failed to enumerate, and jupiter
boot-looped into an emergency shell nobody could use — root is locked, so
sulogin offers a prompt with no answer, and there is no ssh from there:
Timed out waiting for device /dev/disk/by-uuid/dadbff6f-...
Dependency failed for /mnt/data.
Dependency failed for /var/lib/private/prowlarr.
Dependency failed for Local File Systems.
local-fs.target: Job local-fs.target/start failed with result 'dependency'
Reached target Emergency Mode.
`nofail` on /mnt/data did not help, because the prowlarr and seerr bind
mounts layered on top of it had none: without it a mount is RequiredBy
local-fs.target, so those two failed the target on the array's behalf. Give
them `nofail` too and let them fail alone. `systemd.enableEmergencyMode =
false` then keeps a bad array from costing a reachable box at all — far more
useful on a headless host than a console prompt.
Booting further is only safe if nothing quietly relocates onto the 29G eMMC,
so pin the array-backed services to the mount. systemd derives
RequiresMountsFor from a unit's own paths, which for these is somewhere under
/var/lib (eMMC) — nothing pointed immich at mediaLocation or sabnzbd at its
configFile, so with the array gone they would have started and written to the
OS disk, into directories that go invisible the moment /mnt/data mounts over
them. jellyfin, sonarr, radarr and gitea already had a real dependency and
are untouched.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Workflows push as a dedicated ci-bot account rather than a human one, so its
PAT can be scoped, rotated and revoked on its own. Adding a repo to
`ciBotRepos` and redeploying is all it takes to grant access.
Collaborator access and branch-protection push-whitelisting exist only on
gitea's HTTP API — no CLI, no config-file surface — so this one part stays
imperative: a oneshot that PUT/PATCHes the API into the desired state. It
runs on deploys where the script changed, which means it won't self-heal a
revert done through the web UI unless the unit is restarted too.
Two secrets, deliberately distinct:
- gitea_provisioning_token is darman's own token (write:repository +
write:user). Only an owner-scoped token clears reqOwnerCheck on the
collaborator and branch-protection endpoints, and write:user is what lets
it write the Actions secret below. ci-bot cannot grant itself access.
- gitea_ci_bot_token is ci-bot's push token, generated once by hand (the
command is in the comment) and pushed into gitea as a user-level Actions
secret CI_BOT_TOKEN. Gitea has no instance-wide secret scope, and every
repo here is owned by darman directly rather than an org, so a user-level
secret is the closest thing — repo-level lookups fall back to it.
Branch protection is applied to the default branch plus `develop`, since
version-bump.yml pushes there.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
c5a231b pinned the hash by hand because nixpkgs bumped 1.98.8->1.98.9 without
updating vendorHash (NixOS/nixpkgs#545860). The previous commit's lock moves
nixos-26.05 past the point where that fix was promoted from release-26.05, so
the override is now dead weight — and a stale vendorHash override is worse
than none, since it silently wins over a correct upstream value on the next
version bump.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- swap gtk portal/apps for cosmic (xdg-desktop-portal-cosmic, cosmic-files,
cosmic-settings) and drop dolphin/protonplus/bambu-studio (bambu-studio
moved to flatpak alongside the other comms/gaming flatpaks)
- mount jupiter's samba share at /mnt/jupiter (automount, credentials from
the same samba_password secret jupiter itself uses)
- add tea (gitea's remote API CLI) for talking to git.mgaction.town from
terra without SSHing into jupiter
- new dark icon themes (Amy, Azure Glassy, Slot Beauty) vendored from
gnome-look.org tarballs, packaged since pling download links expire
- rishot: fix Qt5Compat.GraphicalEffects QML import (was missing qt6.qt5compat
on QML_IMPORT_PATH, so quickshell failed at config-load)
- launcher widgets: stop LauncherConsole/LauncherDock from reserving
compositor space (ExclusionMode.Ignore, they're overlays not real docks);
bump LauncherCorner app icon size 28->34
- comms script: launch telegram/discord via flatpak, not native binaries
- nix-ld + boot.binfmt aarch64 emulation (for building/flashing mercury
from terra)
TEMPORARY: nixpkgs bumped tailscale 1.98.8->1.98.9 without updating
vendorHash (NixOS/nixpkgs#545860, fixed on release-26.05 but not yet
promoted to the nixos-26.05 channel branch this flake tracks). Remove once
`nix flake lock --update-input nixpkgs` picks up a fixed rev.
Runner registers against the same gitea instance and runs jobs in podman
containers (services/containers.nix), one image per runs-on label using the
catthehacker act-compatible images. Registration token comes from gitea
itself (gitea actions generate-runner-token) and is stored in
secrets/jupiter.yaml, rendered into a TOKEN=... env file via sops.templates
since gitea-actions-runner takes an EnvironmentFile, not a raw secret path.
- Hyprland workspace rules: start-communications.sh launches telegram +
discord into special:communications; qbz/discord/telegram switched to
flatpak (nix-flatpak, Flathub) — removes qbz and proton-pass-cli flake
inputs
- proton-pass-cli and claude-code sourced from nixpkgs-unstable; unstable
pkgs set threaded into home-manager via extraSpecialArgs
- GTK/libadwaita dark theme fixed: dconf color-scheme = prefer-dark written
declaratively instead of a per-session gsettings call
- scripts/keys: store/restore SSH host keys and sops age keys via Proton
Pass (ssh_host#<config> / age#<config> / age#admin naming); no jq dep,
uses pass-cli --field directly
- jq added to desktop-apps system packages
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Hyprland workspace rules: start-communications.sh launches telegram +
discord into special:communications; qbz/discord/telegram switched to
flatpak (nix-flatpak, Flathub) — removes qbz and proton-pass-cli flake
inputs
- proton-pass-cli and claude-code sourced from nixpkgs-unstable; unstable
pkgs set threaded into home-manager via extraSpecialArgs
- GTK/libadwaita dark theme fixed: dconf color-scheme = prefer-dark written
declaratively instead of a per-session gsettings call
- scripts/keys: store/restore SSH host keys and sops age keys via Proton
Pass (ssh_host#<config> / age#<config> / age#admin naming)
Tome (née AudibleLibrary) is darman's own Photino/.NET desktop app, private
repo on our own gitea. Fetched as a flake input over ssh with darman's
ambient key — same mechanism as any other git input, private or not.
buildDotnetModule package: the Preact/Vite frontend (Tome.App/ClientApp)
builds as its own buildNpmPackage derivation and gets copied into the
published app's wwwroot, since upstream's in-project MSBuild npm target has
no network access in the Nix sandbox. Photino.Native's runtime deps (gtk3,
webkitgtk_4_1, libnotify) are wrapped in — confirmed via readelf/ldd that
this Photino build already targets webkit2gtk-4.1, not the now-removed 4.0.
Also added dotnet-sdk + nodejs to terra for developing Tome locally, and
vivaldi (unfree, extends the existing allowUnfreePredicate).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Not in nixpkgs — upstream ships a shell launcher + QML tree with no build
step, driven entirely by `qs -p <dir>`. Packaged as a stdenvNoCC derivation
that wraps the launcher with RISHOT_CONFIG_DIR (sidesteps its argv0-relative
self-lookup, which wrapProgram breaks) and its runtime deps on PATH.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Replaces CachyOS on the OS SSD (Kingston SA400, disko-managed). Dev-data
disks (sdc ext4 /mnt/hdd_01, LVM vg_ssd /mnt/ssd_01) stay out of disko and
are mounted as plain filesystems so they're never wiped. Desktop split into
services/desktop/desktop-hyprland.nix (session: compositor, greeter, audio,
portals) and desktop-apps.nix (things darman actually launches, including
claude-code — allowlisted alongside the other unfree desktop apps).
clientId was still the placeholder "immich" instead of Authentik's actual
generated id, and the sops secret it points at (immich_oauth_client_secret)
was never declared on jupiter. Wire both up, and set endSessionEndpoint so
logout lands back on immich instead of Authentik's "logged out" page.
jupiter had a leftover docker-compose Immich on the RAID (/mnt/data/Immich,
9.9G) that survived the NixOS install. Native module now, media at
/mnt/data/AppData/immich, caddy vhost on 2283 with a 50GB body limit
(caddy's default rejects video uploads).
The package comes from nixpkgs-unstable, the module from the 26.05 pin:
26.05 ships immich 2.7.5, but that database was last written by 3.0.0 and
migrations only run forward --
corrupted migrations: previously executed migration
1776217577402-DropAuditTable is missing
Safe because the two module files are byte-identical at these revisions;
services/media/immich.nix carries the diff command to re-check on a bump.
Drop the input once the stable pin ships >= 3.0.0.
immich needs group "users" only to traverse /mnt/data/AppData (drwx--x---);
its own dir stays 0700 immich:immich. mediaLocation is outside /var/lib, so
the module's tmpfiles entry only ADJUSTS it -- add a rule that creates it.
scripts/immich-import-legacy-db does the database half: boots a copy of the
legacy PGDATA under the matching image (PG14 + vchord 0.3.0 + pgvector
0.8.1), dumps it with the local pg_dump 17, restores into a scratch DB,
fixes ownership, and only swaps after confirmation. Never touches the
original. The old cluster ran VectorChord, not pgvecto.rs, so the smart
search and face embeddings survive -- no ML re-run.
Imported: 666 assets, 25 people, 647 clip + 359 face embeddings, 2 users.
Neither is imported by any host -- both are parked while their upstreams
settle (cinephage ships only a container image, mediamanager comes from a
community flake). Grouping them apart from services/media keeps that
category to what jupiter actually runs.
Pure rename, no content change; nothing imported them, so no host config
moves with them.
Three connected changes, all triggered by the same outage.
base_domain leaves mgaction.town. That zone has a wildcard A+AAAA pointing
at neptun, and DNS wildcards match multi-label names, so
jupiter.hosts.mgaction.town resolved publicly to NEPTUN and Caddy proxied
to itself -- a silent loop rather than a lookup failure. Nesting the
tailnet inside the LAN domain as orbit.sol keeps the theme and resolves
unambiguously, since tailscale matches routes by longest suffix.
override_local_dns = true with pihole as the only global nameserver, so
roaming devices get ad blocking and .sol names off-LAN. With it false,
globalResolvers land in the netmap's FallbackResolvers, which a phone
with carrier DNS never consults. No public fallback is listed on purpose:
tailscale treats the list as a set, so a second entry would let queries
slip past the filter whenever mercury is slow. The cost is that mercury
is now a single point of failure for tailnet DNS.
neptun and mercury opt out individually. mercury would otherwise resolve
through itself. neptun must not depend on a Pi behind a domestic line to
renew the certificates for the control server every other node needs --
and it is circular besides, since tailscaled has to resolve
vpn.mgaction.town to connect at all. Instead neptun runs a dnsmasq stub
forwarding just orbit.sol to MagicDNS on 100.100.100.100, which tailscaled
answers whenever it is running regardless of --accept-dns. That resolves
jupiter live, so the hardcoded /etc/hosts pin is gone.
Also sets dns.nameservers.split explicitly: nixpkgs renders its own
dns.split option one level too high, but headscale reads
dns.nameservers.split (hscontrol/types/config.go:722) and so does
headplane, whose DNS page dies on the missing key with "Cannot convert
undefined or null to object". The module's option is dead as written.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
FTL could not write gravity.db, reporting "attempt to write a readonly
database". The database file was writable; the directory was not. sqlite
creates a sibling gravity.db-journal for every write transaction, so FTL
needs to CREATE files in /var/lib/pihole, and the tmpfiles rule left it
root-owned. The error names the database rather than the directory, which
sends you looking at the file and the filesystem, neither of which is at
fault.
Own the directory as 1000 instead -- the pihole user FTL drops to after
the entrypoint's root phase. Podman is rootful here with no userns
remapping, so the number is the same inside and out; on the host it
collides with darman, harmlessly.
The blocklists are now declared in this module and seeded by a oneshot,
because /var/lib/pihole is not declarative and a reflash took gravity
with it. INSERT OR IGNORE keyed on the URL is idempotent so it can run on
every boot, while the expensive rebuild only runs when gravity is empty.
Adding a list to the Nix attribute needs a manual `pihole -g` -- that is
deliberate, since the rebuild downloads every list and is slow on a Pi.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two changes to the control server, plus the fallout on the hosts.
OIDC via Authentik, so `tailscale up --login-server ...` opens a browser
instead of needing a pre-auth key. This is a second Authentik application,
separate from headplane's, with headscale's own /oidc/callback redirect.
Headless hosts keep using pre-auth keys. Note that headscale runs OIDC
discovery at startup and a failure is FATAL -- pointing `issuer` at an
application that does not exist yet means the control server will not
boot, so verify the discovery document before deploying.
override_local_dns = false, because the upstream default of true replaces
resolv.conf with 100.100.100.100 on every node. That silently broke the
LAN's `.sol` names -- pihole serves those and the global nameservers
return NXDOMAIN for them -- and took ad blocking down with them. It also
made each node's entire DNS depend on tailscaled, which is what had
forced --accept-dns=false onto neptun and mercury individually; both of
those workarounds are now removed, and with MagicDNS resolving properly
again neptun no longer needs its hardcoded /etc/hosts pin for jupiter.
Also serves jellyfin and seerr from jupiter, matching the ports they
already use on its LAN vhosts, and rotates the tailnet pre-auth keys.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Authentik provider and application now exist (slug "headplane", which
is what makes the issuer .../application/o/headplane/), so the client ID
is a real value rather than a placeholder, and the client secret and
headscale API key are in sops.
The tailscale pre-auth keys for neptun and jupiter are rotated because
the tailnet was recreated from scratch: the old headscale database went
with the VPS's OS disk, so every key issued against it is meaningless to
the new control server.
Note the headscale API key defaults to a 90d expiry. When it lapses
headplane stops listing nodes with no obvious cause -- `headscale apikeys
list` shows the date.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
By default headscale fetches https://controlplane.tailscale.com/derpmap/default
at startup and treats failure as fatal, so it cannot boot when that URL is
unreachable. A self-hosted control plane that will not start without
Tailscale's infrastructure rather misses the point of self-hosting -- and
it crash-looped for exactly that reason while neptun had no DNS.
Enable the embedded DERP server on region 999 and drop the upstream map.
The relay rides Caddy on :443, which is why that vhost already sets
flush_interval -1; only STUN needs a port of its own.
Verified against headscale 0.28.0 before committing: it starts clean with
urls = [], registers "DERP region: {RegionID:999 ...}" pointing at
vpn.mgaction.town with DERPPort 443, and brings up STUN.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Probing the live Debian VPS turned up three mismatches between what it
serves and what this config declares:
- git.mgaction.town had no vhost at all. Gitea's web UI and HTTPS clones
are public today; only its SSH side (the :2222 socat forward) had been
ported, so a deploy would have taken the web side offline.
- Audiobookshelf is served as abs.mgaction.town, not the longer
audiobookshelf.mgaction.town this config used. The mobile app is
configured with the short name.
- The apex returns 200 from Caddy. Left unserved deliberately, so it now
gets Caddy's default 404; noted in a comment so it doesn't look like an
oversight next time.
Gitea's ROOT_URL was http:// while Caddy terminates TLS for that name.
Gitea builds absolute URLs from it, so clone buttons, redirects and
webhooks were handing out downgraded links.
Also record that defaultGateway6 is confirmed rather than assumed --
`ip -6 route show default` on the VPS gives "default via fe80::1 dev
eth0 metric 1024 onlink".
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
nixpkgs only carries Zitadel 2.71, which predates the login-v2 split and
cannot take a v3/v4 database (its migrations are forward-only), so the
instance running on the old Debian VPS could never have moved onto it.
authentik-nix ships 2026.5.4 and tracks upstream closely.
The authentik-nix input deliberately does not follow our nixpkgs, per
upstream's warning that overriding it breaks their pinned python
dependency set. That costs a second nixpkgs in the lock, so add
nix-community's Cachix to common.nix -- without it the closure is ~400
local derivations (npm, rust, python). The laptop that runs
scripts/deploy needs the same two lines in /etc/nix/nix.custom.conf.
Authentik's own module creates the database and orders its units against
postgresql.target, and recent versions need no redis, so the wiring is
just the module plus a secret. Pin postgresql explicitly so that editing
system.stateVersion can never silently demand a pg_upgrade of the
identity store.
Secret ownership is not uniform and the difference matters: authentik
and caddy take a systemd EnvironmentFile, which PID 1 reads as root
before dropping privileges, so root:root 0400 is correct. Headplane
opens its secret paths itself while already running as the headscale
user, so those three need an explicit owner or they fail to start.
Also on neptun:
- Pass Caddy's ACME account email through the same EnvironmentFile
mechanism and reference it with the Caddyfile {$VAR} placeholder.
services.caddy.email would render the address into the world-readable
store.
- Stop accepting MagicDNS from our own control server. headscale pushes
override_local_dns, so joining the tailnet would point neptun's
resolv.conf at a MagicDNS served by the tailscaled neptun itself hosts
-- a tailscaled failure would then also take out DNS, ACME renewal and
finally the certs for the control server every other node needs in
order to recover.
- Give headplane a writable DNS extra-records file. Its view of
headscale's config stays read-only, which is the right outcome for a
declarative box; records are data rather than config.
- Require a password for sudo. Deploys become interactive, but darman's
key is otherwise the only thing between the public internet and root.
- Enable zram (8 GB, and disko leaves no room for a swap device), and let
tailscaled-autoconnect retry instead of failing permanently when the
control server isn't up yet on a first boot.
networking.hosts still carries a PLACEHOLDER address for jupiter --
replace it from `headscale nodes list` once jupiter first enrols.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Group service modules by category (media, network, vpn, identity,
dev, desktop) to make the growing services/ dir easier to navigate.
containers.nix stays at the top level since it's a shared backend,
not a single-category service.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Shorten verbose multi-paragraph comments to essentials, and drop a
stale claim in common.nix that jupiter kept its own copy of the base
config (it now imports common.nix directly).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Path-route Headplane under /admin on the same vhost as headscale instead of
its own subdomain - Caddy handle blocks split on the prefix, headscale gets
everything else. base_url drops to the site root since Headplane appends
/admin (and the OIDC callback path) itself.
Wire Zitadel as the OIDC provider. client_id/client_secret/the headscale
API key can't be real until Zitadel and headscale are actually deployed and
an application/key exist, so those are REPLACE_ME placeholders for now
(documented in services/headplane.nix) - direct API-key login stays enabled
as a fallback so this can't lock anyone out in the meantime.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Headscale is the tailnet control server every host's services/tailscale.nix
already points at (--login-server=https://vpn.mgaction.town). MagicDNS
base_domain "hosts.mgaction.town" matches the "jupiter.hosts.mgaction.town"
names already used in this repo's Caddy vhosts.
Headplane is its web UI, running as headscale's own user (native process
integration, no container). No OIDC wired up - log in with a headscale API
key generated on the box. Both proxied through Caddy; headscale's vhost
needs flush_interval -1 since its node-update endpoint is a long-poll.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Local Postgres, peer-authed over the unix socket (the "zitadel" role is
granted createdb+createrole and doubles as both the runtime and bootstrap
DB user - no password anywhere). TLS terminates at Caddy; Zitadel listens
on localhost:8080 and is proxied at auth.mgaction.town.
Master key and admin bootstrap password come from sops - the admin
password specifically needs the sops.templates -> rendered-file route
(services.zitadel.steps would leak it into the world-readable Nix store),
same pattern as mercury's pihole.env.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- sabnzbd, prowlarr, sonarr, radarr, clonarr, seerr, cinephage, mediamanager
services, wired into jupiter with LAN Caddy vhosts.
- Gitea: migrated the old ZimaOS docker instance's data (sqlite db, 4 repos,
no LFS objects) into the NixOS module's default stateDir layout. HTTP via
Caddy; git SSH on its own built-in server at :2222 (not :222 - the unpriv
gitea user can't bind <1024).
- mediamanager-nix flake input for the mediamanager service.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- services/pihole.nix: official pihole/pihole:2026.07.2 via podman, host net,
caps NET_ADMIN/NET_RAW/SYS_NICE/CHOWN; FTLCONF_* env config (upstream unbound,
DHCP 50-200, static lease jupiter, .sol domain, local records)
- unbound: resolveLocalQueries=false (was hijacking resolv.conf to :53 -> boot
DNS deadlock; the real root cause of the earlier failures too)
- password via sops FTLCONF env file; /var/lib/pihole created via tmpfiles
- VM-verified: mercury.sol/jupiter.sol/external all resolve, 0 restarts
mercury is the DHCP server (no lease of its own) so its name wasn't resolvable.
Add explicit dns.hosts A records. VM-verified: both resolve, external unaffected.
- per-host darman_password (distinct hash) in secrets/{jupiter,vps,mercury}.yaml
-> hashedPasswordFile; different console password per host (ssh still key-only)
- mercury: dedicated age key (on boot partition post-flash), sops-nix wired
- AdGuard: module has no secret hook + writable config -> mutableSettings=true,
admin password set via web setup on first boot (never in repo/store)
- services/{samba,avahi,audiobookshelf,containers,caddy,tailscale}.nix
- common.nix grows firewall base + timezone; hosts import what they need
- jupiter/vm/vps import service modules; drop the jupiter/services.nix monolith
- each module opens its own firewall ports; caddy/tailscale shared by hosts
- verified: jupiter/vps/vbox eval + jupiter builds, config equivalent
darmanandClaude Sonnet 5