This repository has been archived on 2026-07-30. You can view files and clone it. You cannot open issues or pull requests or push a commit.
Files
nixos/docs/auto-installer.md
T
beatzaplenty f565e9c2a1
Check NixOS configurations / eval-hosts (pull_request) Successful in 11m6s
Fix auto-install.sh to work standalone, not just baked into the image
Two real bugs, both hit live:

1. Shebang: #!/run/current-system/sw/bin/bash only resolves on an
   already-activated NixOS system -- running the checked-out script
   directly (e.g. from a stock ISO, cloned repo) failed with "cannot
   execute: required file not found" on a non-NixOS box. Switched to
   #!/usr/bin/env bash, which resolves identically on NixOS
   (environment.usrbinenv's own default) and any normal Linux distro.
   Also fixed the file's missing executable bit.

2. FLAKE_BASE_URL: previously depended on pkgs.replaceVars substituting
   a Nix-templated @lanDomain@ placeholder at build time -- meaning it
   only ever worked when baked into the built installer image, not when
   run straight from a checkout (the literal, unexpanded "@lanDomain@"
   string reached git as a bogus hostname). Replaced with LAN_DOMAIN in
   scripts/env.sh (manually kept in sync with variables.nix's lanDomain,
   same pattern as NIX_CACHE_HOST/nixCacheHost already), sourced by the
   script itself like every other script in scripts/. Dropped
   pkgs.replaceVars from modules/installer/common.nix entirely --
   scripts/env.sh is now baked into the image alongside auto-install.sh
   at a matching relative path (/etc/nixos-installer/env.sh next to
   /etc/nixos-installer/installer/auto-install.sh) so the script's own
   relative `source` line resolves the same way in both contexts.

loginShellInit's invocation path and docs/auto-installer.md updated to
match. Verified: shellcheck clean on both scripts, the baked files are
byte-identical to their checked-in sources (no templating left to
verify), and codex-maintenance.sh (secret grep, fmt, statix, full eval
of every host/package including the installer/pxe artifacts) passes
clean.
2026-07-22 02:38:11 +00:00

308 lines
15 KiB
Markdown

# Auto-installer
This flake builds a self-contained NixOS installer environment that can
install any host exposed by its own `nixosConfigurations`. It was migrated
from a formerly-separate `nix-auto-installer` repo — everything it did now
lives here.
The installer provides a small NixOS install environment (ISO, or the same
image netbooted via PXE) with SSH access, Git support, and an interactive
installation script.
Logging in as any user (root or `nixos`) runs
`/etc/nixos-installer/installer/auto-install.sh` (the same file as
`scripts/installer/auto-install.sh` in this repo — see "Installer process"
below for why it's baked in at that path rather than a flat
`/etc/auto-install.sh`), discovers available hosts from this same flake,
lets the operator choose a target, applies that host's Disko storage
configuration, installs NixOS, and reboots.
**This applies to every `nixosConfigurations` target except `lxc-*` hosts —
see "LXC hosts" immediately below for why those are different.**
## LXC hosts
`lxc-*` targets (`lxc-minimal`, `lxc-nix-cache`, `lxc-server`, `lxc-docker`,
`lxc-gui`, `lxc-pxe-boot`, `lxc-tailscale-exit-node`, `lxc-tor-relay`) are **not** installed via `auto-install.sh` — the
interactive menu deliberately excludes them. Don't try to select one there;
`nixos-install` would bind-mount `/` onto `/mnt` (LXC containers have no raw
disk to partition) and then refuse to touch the filesystem it's currently
running on — it's designed to protect exactly this case, so it just fails.
`modules/platforms/lxc.nix` imports nixpkgs' own
`virtualisation/proxmox-lxc.nix` module, which gives every `lxc-*` host a
`config.system.build.tarball` output — a complete, directly Proxmox-importable
container image, no install step at all:
```sh
nix build .#nixosConfigurations.lxc-minimal.config.system.build.tarball
```
This is a plain rootfs tarball, not a `vzdump` backup archive — restoring it
with `pct restore` fails ("archive contains no configuration file"), since
that command expects backup-archive metadata this tarball doesn't have. Use
it as a CT *template* instead: drop it under Proxmox's template storage
(conventionally `/var/lib/vz/template/cache/` for the `local` storage, or
the GUI's "Create CT" → upload-as-template flow) and create a container
from it, supplying all config on the command line since a template has none
of its own:
```sh
pct create <vmid> local:vztmpl/<file>.tar.xz \
--unprivileged 1 --features nesting=1,keyctl=1 \
--rootfs local-lvm:8 --hostname <name> --cores 2 --memory 2048 --swap 2048 \
--net0 name=eth0,bridge=vmbr0,ip=dhcp
pct start <vmid>
```
Every one of those extra flags is load-bearing, confirmed by actually
booting one:
- `--unprivileged 1``modules/platforms/lxc.nix` sets
`proxmoxLXC.privileged = false`, so the image assumes it's running
unprivileged. `pct create`'s own CLI default for this flag is
privileged (unlike the web UI, whose checkbox defaults the other way)
— omit it and you get a privileged container running a NixOS config
that assumes unprivileged, a real mismatch.
- `--features nesting=1,keyctl=1` — required for a modern (v247+)
systemd guest to boot unprivileged at all. Without it, AppArmor denies
the nested user namespaces and credential mounts systemd routinely
uses (even plain getty units) — every getty crash-loops on a denied
`/run/credentials/*` mount every ~3s (this is what garbage on the
console turns out to be) while core services like `nsncd` fail the
same way, and the system never finishes activating.
- `--swap 2048``--memory` doesn't touch swap; it silently stays at
Proxmox's own 512M default otherwise. Match it to `--memory` unless
you deliberately want otherwise.
First boot runs `boot.postBootCommands` (registers the Nix store DB and
system profile) — there's no separate activation step to run yourself.
`scripts/proxmox/create-proxmox-resource.sh --type lxc --host <name>` automates all
of this (host-key handling, building the tarball directly on the Proxmox
node itself, `pct create` with the flags above) — see its `--help`.
Host keys still need pre-seeding the same way as any other host — the
sops-nix activation-vs-first-boot race is identical regardless of how the
image reaches the machine. Unlike the ISO/PXE installer (where
`modules/installer/host-keys.nix` bakes *every* `host-keys/` entry into
`/etc/host-keys/` for `auto-install.sh` to pick from and copy at install
time — see "Host keys" below), an `lxc-*` tarball has no install step to
copy anything during, so `modules/platforms/lxc.nix` bakes this *one*
target's key straight into `/etc/ssh/ssh_host_ed25519_key(.pub)` directly,
keyed by its own exact flake target name (`config.environment.etc` can't
be read back from within a module still contributing to it, so this comes
in via `specialArgs.flakeTarget`, set by `flake.nix`'s `mkTarget`):
```sh
NIXOS_HOST_KEYS_DIR="$(pwd)/host-keys" \
nix build .#nixosConfigurations.lxc-nix-cache.config.system.build.tarball --impure
```
Confirmed the hard way: without this, the tarball's own built-in system
just generates a fresh host key at first boot like any host would, which
can never match whatever `.sops.yaml` actually trusts for that target —
`sops-install-secrets` fails with `Error getting data key: 0 successful
groups required, got 0`, and *every* secret (including this host's own
login) permanently fails to decrypt, silently — no error in the boot log
at all, since the activation step that would install secrets only runs on
a from-scratch first activation and skips silently once `/run/current-system`
already exists. `scripts/proxmox/create-proxmox-resource.sh` always builds with
`NIXOS_HOST_KEYS_DIR` set for this reason.
## Layout
- `modules/installer/common.nix` — shared by every installer target: SSH
access, users, the generated `/etc/auto-install.sh` script, and the
`programs.bash.loginShellInit` hook that runs it on login.
- `modules/installer/iso.nix` — ISO/netboot-specific: imports the stock
`installation-cd-minimal.nix` module plus `common.nix`. Also used, paired
with `netboot-minimal.nix`, to build the PXE netboot variant (see
`docs/pxe-boot.md`).
- `modules/installer/host-keys.nix` — optionally bakes pre-generated SSH
host keys into the image; see "Host keys" below.
- `scripts/secrets/sync-host-keys.sh` — admin-workstation tool that generates,
registers, and (via `--remove`/`--regenerate-all-keys`) retires host
keys; see "Creating a New Machine" below.
- `scripts/secrets/prepare-host-key.sh` — narrower predecessor: generates a single
key by an arbitrary name without touching `.sops.yaml`. Still useful for
pre-generating a key *before* its flake target exists (`sync-host-keys.sh`
can only act on targets `nixosConfigurations` already has); otherwise
`sync-host-keys.sh` does the same thing and more.
Flake outputs:
```nix
nixosConfigurations.installer # ISO/netboot installer image
packages.x86_64-linux.iso # installer ISO/netboot image
packages.x86_64-linux.pxe # netboot-ipxe + netboot-initrd + netboot-kernel, bundled
```
```sh
nix build .#iso
nix build .#pxe
```
There's no `nixosConfigurations.proxmox-lxc` (installer-boots-as-an-LXC-
container) or `packages.x86_64-linux.lxc`/`.all` anymore. Both existed only
to let the installer itself run as an LXC container so you could
`nixos-install` some *other* host from within it — but LXC targets are
excluded from the install menu (same bind-mount problem as any LXC
`nixos-install`), and now have their own direct tarball path anyway (see
"LXC hosts" above), which left the installer's own LXC form with no real
use case.
The `pxe` variant is also built automatically as part of the `pxe-boot` host
itself (`modules/pxe-boot/stage-installer-artifacts.nix`) and served over
iPXE as the menu's "NixOS Auto-Installer" entry — see `docs/pxe-boot.md`.
That same host also builds and serves `packages.x86_64-linux.pxe-minimal`,
a vanilla NixOS minimal netboot image with none of this auto-installer's
wiring, as a separate "NixOS Minimal" menu entry — also documented in
`docs/pxe-boot.md`, not covered further here since it's not this installer.
## Host keys
`sops-nix` derives each host's decryption key from its own
`/etc/ssh/ssh_host_ed25519_key`, generated at **activation** time — before
systemd would otherwise generate one on first boot. Without pre-seeding this
key, secrets (including the root/nixos login password) fail to decrypt on a
genuinely fresh install.
Generated host keys live in `host-keys/` at the repo root (`ssh_host_ed25519_key`
+ `.pub` pairs per hostname). This directory is **gitignored on purpose**
private key material must never be committed — which also means flakes can't
see it through a normal relative path. `modules/installer/host-keys.nix`
reads it through `builtins.getEnv`, which Nix silently returns as an empty
string under normal (non-`--impure`) evaluation, so the module is a no-op —
safe by default, including in CI — unless explicitly opted into:
```sh
NIXOS_HOST_KEYS_DIR="$(pwd)/host-keys" nix build .#iso --impure
```
When built this way, every key currently in `host-keys/` is baked into the
image at `/etc/host-keys/<hostname>_ssh_host_ed25519_key(.pub)`, and
`auto-install.sh` automatically installs whichever one matches the flake
target selected at install time — no manual per-host scp step needed.
**Trade-off, accepted deliberately for this LAN-only setup:** baking keys in
means every key present in `host-keys/` at build time becomes readable by
anyone who can reach the built image — including, for the PXE variant, anyone
who can reach the `pxe-boot` host's unauthenticated HTTP server. This is
considered acceptable here because `pxe-boot` sits behind LAN-only network
infrastructure, not the open internet. If that ever changes, reconsider this
default.
`auto-install.sh` still supports the older manual path as a fallback: if a
host's key isn't baked in (`/etc/host-keys`), it checks `/root/host-keys`
next, where you can `scp` a key in after boot, same as before this migration.
## Storage
Disk partitioning is handled by Disko — the installer has no hardcoded
`parted`/`mkfs`/`mkswap`/`mount` commands, and `auto-install.sh` runs
`disko --mode destroy,format,mount` unconditionally, no branching on whether
the target has a Disko config. Every host reachable through this menu has
one:
- `proxmox-*` (`modules/disko/proxmox.nix`): a real GPT partition table
(ESP + swap + root) on `/dev/sda`.
- `linode-*` (`modules/disko/linode.nix`): Linode provisions and sizes
`/dev/sda`/`/dev/sdb` itself as whole, unpartitioned block devices before
the OS ever boots, so this declares them with `destroy = false` (skips
disko's wipe stage for these disks entirely — see the option's own docs)
and a bare `filesystem`/`swap` content type with no partition table, and
the format step it does run only calls `mkfs`/`mkswap` if `blkid` shows
the device isn't already formatted — a re-run against an
already-provisioned Linode disk is a no-op, not a wipe.
`lxc-*` is the only category without one — it's excluded from this menu
entirely (see "LXC hosts" above), so it never reaches this code path.
## Installer process
`scripts/installer/auto-install.sh` is a real, version-controlled shell
script — not an inline Nix string. It sources `scripts/env.sh` for
`LAN_DOMAIN` itself (same as every other script in `scripts/`), so it
behaves identically whether it's run straight from a git checkout (e.g.
manually, from a stock NixOS ISO that isn't this repo's own installer
image) or from inside the built installer image. That's also why it's
baked in at `/etc/nixos-installer/installer/auto-install.sh` rather than a
flat `/etc/auto-install.sh``modules/installer/common.nix` bakes
`scripts/env.sh` in alongside it at `/etc/nixos-installer/env.sh`,
preserving the same relative layout (`installer/auto-install.sh` ->
`../env.sh`) the checked-out repo has, so the script's own
`source ".../env.sh"` line resolves correctly in both places without any
Nix-level templating.
Once running, it:
1. Queries `nixosConfigurations` from this flake over the network (`git+https://<lanDomain>/beatzaplenty/nixos.git`) — this happens at *install* time, not build time, so a generic installer image always sees whatever hosts are currently committed, without needing a rebuild.
2. Presents them as a menu; confirms the choice.
3. Skips the `nix-cache` substituter when installing a `nix-cache` host itself (consistent with that host's own runtime config).
4. Runs `disko --mode destroy,format,mount` (see "Storage" above — every host reachable through this menu has a Disko config, so this is unconditional).
5. Installs the target's SSH host key from `/etc/host-keys` or `/root/host-keys` (see "Host keys" above).
6. Runs `nixos-install --flake <url>#<choice> --no-root-password`.
7. Cleans up and reboots.
## Creating a new machine
Do this instead of jumping straight to a plain install whenever the target
host consumes any sops-nix secret — as of this writing, that's every host
(`modules/common/configuration.nix` puts the root/nixos password hash and the
GitHub token behind sops-nix for all of them).
1. **Add the flake target**`hosts/<name>/host.nix` plus the matching
`mkTarget { ... }` entry in `flake.nix`'s `generatedTargets` (see
"Composition pattern" in `CLAUDE.md`). No secrets involved yet, so this
is safe to commit on its own if you want a clean history.
2. **On your admin workstation, generate and register its host key:**
```sh
./scripts/secrets/sync-host-keys.sh <flake-target>
```
This generates `host-keys/<flake-target>_ssh_host_ed25519_key(.pub)`,
adds it as a new `.sops.yaml` anchor, works out which `secrets/*.yaml`
files this specific host actually references (from its own
`config.sops.secrets`, not guessed), adds it to each one's
`key_groups`, and re-encrypts them with `sops updatekeys` — no manual
YAML editing. Safe to re-run; it only fills in what's missing.
Doing this for every host that needs one at once — after adding several
new targets, or just to catch up any that were missed — is
`./scripts/secrets/sync-host-keys.sh --all`. See `scripts/secrets/sync-host-keys.sh --help`
for its other modes (`--remove`, `--regenerate-all-keys`).
3. **Commit and push.** The flake build the installer uses has to see the
new recipient before you install, or decryption fails on first boot
regardless of the next step.
4. **Build the installer image with keys baked in** (or reuse an already-serving `pxe-boot` host, which does this automatically once redeployed):
```sh
NIXOS_HOST_KEYS_DIR="$(pwd)/host-keys" nix build .#iso --impure
```
5. **Boot it on the target machine**, log in, select the new host's flake
target from the menu, confirm. `auto-install.sh` finds the baked-in key,
runs Disko + `nixos-install`, and reboots.
6. **Verify after reboot:**
```sh
ssh <new-host> ls /run/secrets/
```
If that's empty or login fails, the host's age key most likely wasn't in
`.sops.yaml` (or wasn't re-encrypted into the secrets file it needs) when
`nixos-install` ran — fix `.sops.yaml`/`secrets/*.yaml`, push, then re-run
`nixos-install --flake .#<hostname> --no-root-password` from a rescue
environment against the existing `/mnt`, or just redo the install.
## Safety
This installer is destructive: `disko --mode destroy,format,mount` erases
any disk defined by the selected host's Disko configuration. Always verify
the selected host profile and target machine before confirming.