- Nix 73.5%
- Go 25.2%
- QML 1.3%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
|
All checks were successful
Check / check (push) Successful in 7m27s
A dock icon now gets a tile behind it while the pointer is over it, in the inactive window border colour. While a button is held the tile turns the active border colour and the icon shrinks to 92%. That is the size Hyprland's "popin 92%" window animation starts from, and the number is read from that setting. The tile fills the icon's box with the window corner radius: the icons sit as far inside the dock's inner edge as windows sit inside the main pill, so the tile corners run parallel to the dock's the way window corners run parallel to the pill's. With Papirus, round and folder icons get a full tile and foot's square icon gets a thin ring. Hover, press and the dock slide all use the slide duration and curve, now one Ease component (a PropertyAnimation, so it animates both colour and numbers) in place of the inline NumberAnimation. The dock is now a child of the reveal zone. In Qt 6.12, once QQuickDeliveryAgentPrivate::deliverHoverEventRecursive finds a hovered item with a HoverHandler, it passes hover only to that item's ancestors, never to siblings beneath it. With the icons' new HoverHandlers, the zone, a sibling of the dock, lost hover the moment the pointer reached an icon, and the dock hid under it. Nested in the zone, the dock's icons have the zone as an ancestor. dock.box maps the dock back into the bar's coordinates for the frame paths. Tested 2026-11-16 in a nested Hyprland 0.57.2 with quickshell 0.3.2 and two foot windows. Hovering LibreWolf, foot and Nautilus in turn moved the tile from icon to icon while the dock stayed up. Two clicks on foot alternated focus between its windows, and the dock hid once the cursor left. The windows stayed 774x836. The pointer was moved with wlrctl's relative moves, because a hyprctl warp within one surface sends no motion. wlrctl cannot hold a button, so the pressed look was checked by forcing Steam's pressed state in a test copy. A real held press was not tried. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> |
||
| .forgejo | ||
| docs | ||
| flake | ||
| hosts | ||
| lib | ||
| modules | ||
| overlays | ||
| packages | ||
| roles | ||
| secrets | ||
| services | ||
| sites | ||
| tests | ||
| .betterleaks.toml | ||
| .gitignore | ||
| .golangci.yml | ||
| .sops.yaml | ||
| AGENTS.md | ||
| CLAUDE.md | ||
| flake.lock | ||
| flake.nix | ||
| README.md | ||
| statix.toml | ||
nix-desktop
One NixOS flake for the whole fleet: its services, an AmneziaWG mesh,
self-hosted DNS, sops secrets split by trust boundary, and deploys that roll an
unhealthy host back on their own. How to change it is in
AGENTS.md.
| Host | Kind | Runs |
|---|---|---|
thor |
workstation | Hyprland, work VPN, gaming |
loki |
laptop | Hyprland, work VPN |
heimdall |
home server, mesh hub | media, Forgejo and CI, Matrix, observability, backups |
vidar |
VPS, reached over VLESS/REALITY | DNS, DoH, TURN, mail, backup target |
bifrost |
VPS | AmneziaWG gateway |
magni |
Anbernic RG353V | AArch64 SD image, RetroArch |
forseti |
dron's second machine |
home PC, headless NixOS on a Ryzen 3 2200G; backup target for caco-zero-three, outside the mesh |
caco-zero-three |
tenant-owned GPU server | LLM stack, Hermes, Matrix, Kanidm, UPS shutdown; deployed separately |
modi |
Pixel 9 Pro XL | native AArch64 NixOS, Phosh, Btrfs on userdata |
Each host starts at hosts/<host>/default.nix, which imports its composition.nix.
Layout
flake.nix inputs and the list of flake-parts modules
flake/ one module per output kind: hosts, packages, checks, dev
hosts/ inventory.nix and one directory per host
sites/ operator and tenant identity: user, keys, domains, Wi-Fi
modules/ base.nix, shared options and behaviour, web service registry
roles/ capabilities a host composes
services/ one directory per service
lib/ fleet construction, typed inventory, mesh topology
packages/ flake packages; the Go tools are in command-suite
tests/ checks, registered in flake/checks.nix
secrets/ sops files
How a host is built
Fleet construction supplies modules/base.nix and hosts/<host>/default.nix.
The host imports composition.nix, a normal NixOS module listing its roles
and services. Host-specific service modules import the shared service beside
their overrides. Check a reorder with nix run .#nix-eval-and-diff -- --contents <host>.
hosts/inventory.nix holds only mesh membership and deployment.
lib/inventory.nix types it and checks the fleet-wide invariants, and
lib/topology.nix derives the mesh subnet, peer lists and internal DNS from it.
Hosts read the results as config.nd.fleet and config.nd.topology. Every
inventory host is a node of one Colmena hive, and nixosConfigurations is that
hive's nodes.
Roles
Roles import the roles they build on. Put a package in the narrowest one;
tests/software-stack.nix enforces the boundaries.
| Role | Adds | Imports |
|---|---|---|
cli |
fish, Git, small shell tools | |
workstation |
credentials, TPM-backed SSH and signing | cli |
development |
compilers, language servers, Podman | workstation |
operations |
AWG, DNS, SOPS and storage tools | workstation |
ai |
coding agents from llm-agents.nix, and its binary cache | workstation |
desktop |
Hyprland session and default apps | workstation |
work |
mail, VPN, employer certificates | workstation |
gaming |
Steam, emulators, image-quality layers | desktop |
server |
bash, headless diagnostics, server health checks | |
reverse-proxy |
Caddy, ACME, web health checks and hardening |
Services
A service is enabled by listing it in a host's composition.nix, so it has no
enable flag of its own. To find out whether another service is present, a
module asks that service's upstream option, such as
config.services.forgejo.enable.
A host serving websites lists roles/reverse-proxy in its composition alongside
roles/server. Gateways that do not serve websites need only server.
A service behind Caddy declares itself once, and its vhost, hardening floor,
deploy health check and dashboard follow. The fields are documented in
modules/web/options.nix:
nd.services.web.jellyfin = {
port = 8096;
unit = "jellyfin";
job = "jellyfin";
health = "/health";
probed = true;
hardening = "runtime";
blockMetrics = true;
};
With required = true, a failing health check rolls the deploy back. Set it
only for services whose loss costs access to the host or the ability to deploy,
such as Forgejo. gate = { }; puts the vhost behind a Kanidm login, and
gate.except or gate.only exempt paths whose clients cannot follow an OAuth
redirect.
A service with state worth keeping lists that state in
nd.backups.state.<name>.paths. Each contract selects a backup job, defaulting
to nd.backups.defaultJob. The host imports services/backups and declares
nd.backups.jobs.<name> with runtime repositoryFile and passwordFile paths,
an optional environmentFile, and any unit dependencies needed to reach it.
The job derives its paths, quiesced units, integrity checks and restore checks
from its assigned contracts. A declaration without a consuming job fails
evaluation, even when the backup service is absent. An explicit excludeReason
instead reports a warning and makes no claim of backup coverage. Current
exclusions and the steps to close them are in
docs/flake-contracts-2026-10-27.md.
caco-zero-three's destination is forseti's Western Digital disk, reached over
the home LAN through a tenant-owned restic repository whose credentials live in
each end's own host secrets file.
OAuth integrations declare nd.services.oauthClients.<client> beside the
application, with its callbacks, landing URL, scope maps and secret reference.
services/kanidm consumes the registry. Confidential clients get a runtime
SOPS secret from this host's encrypted file; public clients require PKCE and
have no secret. Web gates declare their clients through the same registry.
Caco's checkpoint manifests in hosts/caco-zero-three/llm/profiles.nix pin
repository revisions and each file's byte size and SHA-256. The downloader
checks installed files, resumes downloads into .part files, and renames a
download into place only after its hash matches. The projector can come from
a different repository while retaining its existing model-directory path.
Deploying
nix run .#deploy -- <host|all> [build|dry-activate|health|switch] [--reboot]
nix run .#deploy-caco -- all switch
nix run .#rebuild-current # the machine you are on, through nh
switch is the default goal. It runs the current generation's own health
probes first, then builds and dry-activates the whole selection, then switches
the hosts one at a time. Before each activation it arms a timer on the target
that restores the captured generation. The timer is cancelled only after
/run/current-system is the expected closure and that closure's own probes
pass. A host that loses SSH partway through therefore rolls itself back.
A generation that changes systemd or the D-Bus implementation cannot switch
live. --reboot installs it with boot, reboots the host, and rolls back if
the host comes up unhealthy. --allow-busy skips the busy check (active
LiveKit rooms), and unsafe-switch skips every check. To give a slow host more
time, raise DEPLOY_ROLLBACK_TIMEOUT_SEC, DEPLOY_HEALTH_ATTEMPTS,
DEPLOY_HEALTH_INTERVAL_SEC or DEPLOY_REBOOT_TIMEOUT_SEC. The code is in
packages/command-suite and is tested by tests/deploy-safe-switch.nix.
deployment.tenant splits the hosts by owner. A tenant's hosts can be deployed
only with that tenant's own command, so deploy all never reaches caco.
CI
CI runs on Forgejo on heimdall. Check (flake-check.yaml) runs lint and then
fleet-check on every push, inside the forgejo-ci container; a newer push
cancels an older run. Deploy and Deploy caco-zero-three are started by hand
from the Actions tab. They run on heimdall's nix-deploy runner and refuse a
commit whose newest Check run did not succeed. thor's
fleet-check-aarch64.timer builds the ARM checks weekly under QEMU.
Each runner registers with a uuid/token pair. Mint the pair on the repository, never on the instance, because an instance runner takes jobs from every repository:
curl --fail https://git.nestoris.online/api/v1/repos/npilosov/nix-desktop/actions/runners \
-H "Authorization: token $FORGEJO_TOKEN" \
--json '{"name":"heimdall-nix","ephemeral":false}'
The uuid goes in services/forgejo/runner.nix. The token goes in
secrets/hosts/heimdall.yaml as forgejo-runner-nix-token, or
forgejo-runner-deploy-token for heimdall-deploy. The deploy SSH key is
forgejo-deploy-ssh-key in the same file, and its public half is in
sites/deploy-authorized-key.nix.
Secrets
Each host decrypts with its own SSH host key, and there is no shared age key.
.sops.yaml lists the recipients of each file.
| File | Readable by | Holds |
|---|---|---|
secrets/common.yaml |
the fleet, not caco | password hash, DoH token, AWG gateway key |
secrets/mesh.yaml |
every mesh member | the mesh's AWG header key |
secrets/desktop.yaml |
thor, loki | Wi-Fi |
secrets/work.yaml |
thor, loki | work, mail, GitLab, Forgejo and Jira credentials |
secrets/cluster.yaml |
heimdall, vidar | backup and TURN credentials |
secrets/turn-caco.yaml |
heimdall, caco | caco's TURN secret |
secrets/hosts/<host>.yaml |
that host | its own keys and service credentials |
To re-key a host after its SSH key changes, convert the new public key with
ssh-to-age, replace the host's recipient in .sops.yaml, and run
sops updatekeys on every file the host reads.
DNS
systemd-resolved -> dnsproxy -> Caddy DoH on vidar -> Blocky -> Knot Resolver
vidar queries its local Blocky directly. The work VPN's DNS servers take
precedence for office domains because of their routing domains. On the home
LAN, hosts find each other over resolved's mDNS on
nd.host.network.mdnsInterfaces, and no host runs Avahi.
New host
- Write
hosts/<host>/default.nix. It imports./hardware.nix(fromnixos-generate-config --show-hardware-config) and the host'ssites/module, if it has one, and it pinsnd.meta.stateVersion.{system,home}. - Write
hosts/<host>/composition.nixwithimports, and import it from the host'sdefault.nix. - Add the host to
hosts/inventory.nix. Setdeployment.enable = trueif it should be deployable. - If the host has a
disks.nix, runnix run github:nix-community/disko -- --mode disko hosts/<host>/disks.nix. Then install withnixos-install --flake .#<host>.
Recovery
nix build .#installer builds an ISO with SSH, iwd, disko and git. If these
secrets are decrypted first, the ISO joins the mesh as bootstrap
(10.90.100.8) and connects to the office Wi-Fi. The secrets are embedded in the
image, so delete it after use.
sops decrypt --extract '["mesh-private-key"]' secrets/hosts/bootstrap.yaml > /tmp/nix-bootstrap-mesh.key
sops decrypt --extract '["awg-header-protection-key-mesh"]' secrets/mesh.yaml > /tmp/nix-bootstrap-mesh-header.key
sops decrypt --extract '["eap-password"]' secrets/work.yaml > /tmp/nix-bootstrap-wifi-eap.key
nix build .#installer --impure
nd.meta.bootstrap = true turns off everything that needs secrets. Use it
until the host has its sops identity back, either by restoring its old
/etc/ssh/ssh_host_ed25519_key or by re-keying it as described under Secrets.
modi
Its first switch has to run on the phone. A fresh nix-on-droid ships a proot that is too old for this flake's glibc, so stage the pinned one from inside the app first:
p=/nix/store/dvf2ck9bkw7yyrlkjk87xz1anaxsgrd6-proot-termux-static-aarch64-unknown-linux-android-unstable-2026-02-20
nix-store -r "$p" && cp "$p/bin/proot-static" /bin/.proot-static.new
Restart the app, then run nix-on-droid switch --flake <checkout>#modi. This
takes about half an hour. After that, build on thor with
nix run .#modi-switch -- <phone address>; the phone's sshd listens on port
8022. Over USB, run adb forward tcp:8022 tcp:8022 and use 127.0.0.1. Over
Wi-Fi, exclude com.termux.nix from the phone's VPN tunnel.
Pixel phone
modi is the native Pixel 9 Pro XL host. Its reproducible kernel, boot images,
firmware provisioning and Btrfs installation are documented in
docs/modi.md.