The installer ISO no longer needs a prebuilt system: `router-setup` asks
for the disk, the WAN and LAN interfaces, DHCP or PPPoE (optionally on a
VLAN), host name, LAN address and a password, writes settings.json and
the secrets, builds that router offline (the ISO carries a reference
router system whose closure has everything the build needs) and installs
it with `router-install --system`.
The password is used for the web UI and, hashed, as root's console
password (routerSettings sets root's hashedPasswordFile; it applies when
the account is created at installation).
checks.setup answers router-setup on the console and checks the installed
box: its settings, the web UI and console login with that password, and
a rebuild that reproduces the installed system.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AeFqV9dZrK6nipHkPfyM4F
nixosModules.settings builds a router from a JSON file: the router
options as they are, plus host name, time zone, SSH keys and
stateVersion; the module system validates it. lib.mkRouterSystem builds
such a system, lib.mkConfigDir the configuration directory installed to
/var/lib/router/config (a flake pinned to this router source, the
settings and an optional local.nix). router-apply builds it and switches
with commit-confirm.
The device must be able to rebuild without internet, so the packages the
rebuilt derivations need are part of the system; they were found by
diffing the derivation graphs of typical changes against the installed
system.
checks.apply installs such a router, boots it offline, checks that the
unchanged settings rebuild to the installed system, applies and
confirms a change, and rolls back an unconfirmed one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
nixosModules.image builds a GPT disk image with systemd-repart: ESP with a
UKI carrying the store's verity root hash, an erofs /nix/store on a
dm-verity partition, a tmpfs root, an immutable /etc overlay and an ext4
state partition on /var. No Nix on the device; it is replaced as a whole.
The store is made with 4 KiB erofs blocks: systemd-repart passes the 512
byte sector size to mkfs.erofs, and libblkid cannot verify the checksum of
such a superblock, so udev never reported the file system and the initrd
waited forever for /dev/mapper/usr.
examples/image (nix build .#image) is a hardware test build with a debug
console password; checks.image boots the image under UEFI.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
DNS and DHCP (needed to replace a dnsmasq-based router):
- router.dns.domain/hosts/overrides/forwardZones: authoritative local
domain, split-horizon host records with reverse entries, wildcard
overrides and per-zone forwarders (private zones skip DNSSEC).
- lans.<n>.dhcp.reservations: Kea host reservations, published in DNS.
- Tunnels with dns = true admit what their peers route, so a
site-to-site peer without a transfer address can use the resolver.
Installer:
- router-install wipes a disk, creates an ESP and an ext4 root and
installs a prebuilt system with its secrets; nothing is built on the
box. installed-layout is the shared disk layout contract.
- lib.mkInstaller builds an ISO for a given nixosSystem;
packages.installer-example is the example's ISO.
Covered by the dns and installer VM tests (the latter installs onto an
empty disk and boots it under UEFI), wireguard DNS and assertion cases.
PLAN.md records the remaining gaps for replacing the hand-written router.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
- Allow the WAN link to also carry LANs when the uplink is PPPoE on a
VLAN with IPv6 off; only the PPP device counts as an uplink then, so
LAN traffic can never match WAN rules.
- Add router.rollback: router-arm / router-confirm / router-rollback
plus `nix run .#deploy`, a commit-confirm deployment flow.
- Wait for the LAN devices instead of the uplink in networkd
wait-online, so an unplugged uplink cannot block boot or activation.
- Add examples/desktop (single NIC, PPPoE on VLAN 1001) as
nixosConfigurations.example, built as a flake check.
Covered by the onestick and rollback VM tests and new assertion cases.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
router.wireguard.<name> configures tunnels through networkd: road warrior
(listening, dual-stack), site-to-site (no transfer network) and provider
dial-out with masquerade. A tunnel has a zone like a LAN, so inter-zone rules
apply; management, DNS and internet are opt-in. Peers' allowedIPs become
routes (destination-based routing; default-route tunnels are rejected).
Tunnels have a kill switch (default on): forwarded packets for their
allowedIPs are dropped if they would leave via an uplink, before the
established-accept so flows that used the tunnel are cut too.
Private keys stay outside the Nix store and are passed to the unprivileged
systemd-networkd with LoadCredential, so key files can be root-only.
Assertions cover CIDR/key/endpoint formats, duplicate allowedIPs, overlaps
with LAN subnets, port and device clashes, and rule destinations inside
tunnel networks. A five-VM `wireguard` check covers road warrior, site-to-site,
the provider tunnel with an IPv6 transfer net, zone rules and the kill switch.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
LANs can sit on an 802.1Q VLAN (`vlan` on a trunk `interface`); networkd
creates the device and brings the trunk up. LANs belong to a named `zone`
(default: their own name). Traffic inside a zone is forwarded freely, traffic
between zones is denied unless a `router.firewall.rules` entry allows it
(protocol, ports, optional destination inside the target zone).
Internet access is now opt-in per LAN (`internet`, default false), replacing
the previous allow-out default. Assertions reject unknown zones, bad
destinations, duplicate devices, over-long device names and a WAN trunk.
Kea subnet IDs now come from the network address so adding LANs does not
renumber them and orphan persistent leases.
Tests: a five-VM `vlans` check (untagged + two VLANs, shared zone, one-way
rules, per-host destination restriction), new assertion cases, parseNet unit
tests, and shared test tooling in tests/common.nix.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
IPv6: router.wan.ipv6.enable mirrors the existing uplink config (DHCPv6 with
WithoutRA=solicit, delegated prefix). The ISP delegates a single /64, so
exactly one LAN may claim it via router.lans.<name>.ipv6.pd (asserted); other
LANs stay IPv4-only. SLAAC only, no NAT66.
PPPoE: router.wan.ipv4.type = "pppoe" runs pppd on an optional VLAN of the
uplink, with the password read from a file at service start (never the Nix
store). NAT, forwarding and MSS clamping (MTU 1492) use the derived PPP
device; native IPv6 stays on the untagged link.
Tests: basic.nix is parameterised by uplink type; the pppoe variant uses a
fake ISP with pppoe-server (VLAN 7, CHAP). Both variants cover IPv6 PD, SLAAC,
routing without NAT and WAN-deny. Assertion tests gain PD/PPPoE cases and a
full-system evaluation step.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Add an IPv4 helper lib (parse, netmask, cidr, overlap) and derive
router.lans.<name>.ipv4.subnet from address and prefixLength instead of
requiring it twice. New assertions reject invalid addresses, DHCP pools that
are reversed, outside the subnet or contain the router address, and
overlapping LANs.
Add `lib` unit tests and an `assertions` check that evaluates broken configs
and compares the exact error messages.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
router.* schema (wan, lans, dns.forwarders) compiled to systemd-networkd,
a generated nftables ruleset with NAT, Kea DHCPv4 and Unbound, plus build-time
assertions against lock-out and interface conflicts.
tests/basic.nix is a three-node VM test (upstream/router/client) covering WAN
DHCP, LAN DHCP and DNS, NAT, WAN-deny and LAN SSH/ping. `nix run .#lab` boots
the same topology in tmux with one console pane per VM and the driver REPL.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
nilsandClaude Opus 5.5