add agent-lab sandbox VM on muede-pc2

This commit is contained in:
müde 2026-11-18 18:40:12 +01:00
commit 873622dc8a
6 changed files with 324 additions and 0 deletions

View file

@ -44,3 +44,21 @@ Machines are configured to act as build servers / binary caches for each other i
```
4. Rebuild all machines so they trust the new signing key.
## agent-lab sandbox VM
`agent-lab` is a throwaway NixOS VM on muede-pc2 for hyperhive agents to ssh
into and apply risky configs to. The host (`nixosConfigurations/muede-pc2/agent-lab-vm.nix`)
only seeds its disk and runs qemu — it never declares the guest's system, so
the guest can `nixos-rebuild switch` itself freely.
- Reachable at `muede-pc2:2222` (qemu user-mode NAT forward), root ssh only,
authorized for the lpt2 and pc2 keys plus muede-pc2's build key.
- `sudo agent-lab-reset [flake-ref]` rebuilds the seed image, replaces the
disk, and restarts the VM. Takes seconds once the image is in the store.
Defaults to the upstream flake; pass a local path to test uncommitted work.
- `/etc/nix/netrc` is copied in over ssh on every VM start by
`agent-lab-vm-netrc.service`, so the private cache credential stays out of
the nix store. The host needs `/etc/nix/distributed-build-key` for this.
- The guest's own config is `nixosConfigurations/agent-lab/`; it is the seed
only. Whatever the agents switch to afterwards is not tracked here.

View file

@ -15,6 +15,10 @@ in
# system = "aarch64-linux";
# nixosSystem = nixos-raspberrypi.lib.nixosInstaller;
# };
agent-lab = {
# sandbox VM on muede-pc2; see nixosConfigurations/muede-pc2/agent-lab-vm.nix
system = "x86_64-linux";
};
compute1-vm = {
system = "x86_64-linux";
};

View file

@ -0,0 +1,39 @@
{
imports = [ ./hardware.nix ];
config = {
# A deliberately disposable box: hyperhive agents ssh in as root and apply
# whatever config they are testing, up to and including breaking it. The
# host (see ../muede-pc2/agent-lab-vm.nix) only ever seeds the disk and
# runs qemu — it never declares this system's config, so the guest is free
# to `nixos-rebuild switch` itself out from under the seed.
# Agents log in by certificate, not by key. sshd accepts one the swarm's
# SSH CA signed, and the single principal on it names the agent, so the
# list below is what admits atlas and maia and nobody else. Dropping a name
# ends that agent's logins at its next attempt; a signed certificate
# expires within three hours either way.
#
# Spelled out rather than imported from hyperhive.nixosModules.agent-ssh-target
# on purpose: this VM should not take the hyperhive flake as an input.
# Re-read the CA key from the swarm's store after the CA is regenerated:
# bao read -field=public_key ssh-agents/config/ca
environment.etc."ssh/hyperhive-agent-ssh-ca.pub".text =
"ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAILHGXr3Voq/SVorEoBY5cDMqTumQyHMweNQvWcSvTXuN\n";
services.openssh.settings.TrustedUserCAKeys = "/etc/ssh/hyperhive-agent-ssh-ca.pub";
users.users.root.openssh.authorizedPrincipals = [
"atlas"
"maia"
];
users.users.root.openssh.authorizedKeys.keys = [
"ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIO1CRn4yYTL4XUdCebE8Z4ZeuMujBjorTdWifg911EOv pc2 home roaming"
"ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIPDNpLDmctyqGpow/ElQvdhY4BLBPS/sigDJ1QEcC7wC lpt2-roaming"
# muede-pc2's build key, reused as the host's provisioning identity: it is
# how /etc/nix/netrc reaches the guest after boot without the credential
# ever passing through the nix store. Private half stays on the host in
# /etc/nix/distributed-build-key.
"ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIHmnyhP6L+kGHV15cb/d31AQr50wSEaQhkUBwy2+OEKk muede-pc2-nix-builds"
];
};
}

View file

@ -0,0 +1,46 @@
{
lib,
modulesPath,
...
}:
{
imports = [ (modulesPath + "/profiles/qemu-guest.nix") ];
config = {
# Legacy BIOS boot via GRUB on the virtio disk, matching the `qemu` image
# variant (`image.efiSupport = false`). The repo-wide systemd-boot default
# would need an ESP plus EFI firmware and a writable varstore in the
# runner, and a guest that reinstalls its own bootloader on every
# `nixos-rebuild switch` is one less thing for an agent to brick. Drop that
# module wholesale -- it also owns boot.loader.timeout and the EFI vars.
my.systemdBoot.enable = lib.mkForce false;
boot = {
loader = {
grub = {
enable = true;
devices = [ "/dev/vda" ];
};
timeout = 1;
};
# The runner resizes the seeded image before first boot (the image itself
# is sized to its closure), so / has to grow into the rest of the disk.
growPartition = true;
kernelParams = [ "console=ttyS0" ];
tmp.cleanOnBoot = true;
};
# Labels, not UUIDs: every reseed makes a brand new filesystem, and the
# point of this machine is being thrown away and remade.
fileSystems."/" = {
device = "/dev/disk/by-label/nixos";
autoResize = true;
fsType = "ext4";
};
# qemu user-mode networking hands out a lease on the single virtio NIC.
networking.useDHCP = lib.mkDefault true;
};
}

View file

@ -0,0 +1,216 @@
{
lib,
pkgs,
...
}:
let
vm = "agent-lab";
unit = "${vm}-vm";
stateDir = "/var/lib/${unit}";
disk = "${stateDir}/disk.qcow2";
qmp = "/run/${unit}/qmp.sock";
sshPort = 2222;
cpus = 4;
memoryMiB = 8192;
diskSize = "60G";
# Deliberately a flake *reference*, resolved when the reset script runs, not
# a store path baked into this host's closure. The host must not pin the
# guest's config: agents rebuild the guest from inside, and a pinned
# `system.build.toplevel` here (the microvm.nix model) would fight that. All
# the host ever does is lay down a first-boot disk.
seedFlake = "git+https://forge.darkest.space/mara/nixos-configs.git";
seedAttr = "nixosConfigurations.${vm}.config.system.build.images.qemu";
# Credentials the guest needs but the nix store must never see. Both are
# hand-placed, out-of-store files on this host (same as on every other
# machine here), and are copied in over ssh after the guest is up.
provisionKey = "/etc/nix/distributed-build-key";
netrc = "/etc/nix/netrc";
# ssh to the guest as the provisioning identity. The host key changes on
# every reseed, and this target is 127.0.0.1 behind qemu's user-mode NAT, so
# host key checking buys nothing here.
guestSsh = lib.concatStringsSep " " [
(lib.getExe' pkgs.openssh "ssh")
"-i ${provisionKey}"
"-o IdentitiesOnly=yes"
"-o StrictHostKeyChecking=no"
"-o UserKnownHostsFile=/dev/null"
"-o LogLevel=ERROR"
"-o BatchMode=yes"
"-o ConnectTimeout=5"
"-p ${toString sshPort}"
"root@127.0.0.1"
];
reset = pkgs.writeShellApplication {
name = "${vm}-reset";
runtimeInputs = [
pkgs.qemu_kvm
pkgs.findutils
pkgs.coreutils
pkgs.systemd
pkgs.nix
];
text = ''
if [ "$(id -u)" != 0 ]; then
echo "${vm}-reset: needs root (writes ${stateDir})" >&2
exit 1
fi
flake="''${1:-${seedFlake}}"
echo "${vm}-reset: building seed image from $flake#${seedAttr}" >&2
out=$(nix build --no-link --print-out-paths "$flake#${seedAttr}")
img=$(find "$out" -maxdepth 1 -name '*.qcow2' -print -quit)
if [ -z "$img" ]; then
echo "${vm}-reset: no qcow2 in $out" >&2
exit 1
fi
systemctl stop ${unit}.service || true
# The tmpfiles rule above makes this exist at activation; keep the
# script able to stand alone if the directory was removed by hand.
install -d -m0700 -o ${unit} -g ${unit} ${stateDir}
# `convert`, not a copy or an overlay: the running disk must not keep a
# backing-file reference into the nix store, or a later GC of the seed
# would pull the VM's filesystem out from under it.
rm -f ${disk}.new
qemu-img convert -O qcow2 "$img" ${disk}.new
qemu-img resize ${disk}.new ${diskSize}
chown ${unit}:${unit} ${disk}.new
chmod 0600 ${disk}.new
mv -f ${disk}.new ${disk}
systemctl start ${unit}.service
echo "${vm}-reset: fresh disk seeded, VM starting — ssh -p ${toString sshPort} root@<host>" >&2
'';
};
in
{
config = {
users.users.${unit} = {
isSystemUser = true;
group = unit;
home = stateDir;
description = "agent-lab sandbox VM";
};
users.groups.${unit} = { };
# `StateDirectory=` in the unit below only creates this when the unit
# starts, and the unit cannot start until the disk exists -- so the very
# first `agent-lab-reset` would have nowhere to write. Declare it here.
systemd.tmpfiles.rules = [ "d ${stateDir} 0700 ${unit} ${unit} -" ];
systemd.services.${unit} = {
description = "agent-lab sandbox VM (ssh target for hyperhive agents)";
wantedBy = [ "multi-user.target" ];
after = [ "network.target" ];
# No disk yet means the VM was never seeded; `agent-lab-reset` makes one.
unitConfig.ConditionPathExists = disk;
serviceConfig = {
User = unit;
Group = unit;
SupplementaryGroups = [ "kvm" ];
StateDirectory = unit;
StateDirectoryMode = "0700";
RuntimeDirectory = unit;
ExecStart = lib.concatStringsSep " " [
(lib.getExe' pkgs.qemu_kvm "qemu-system-x86_64")
"-name ${vm}"
"-machine q35,accel=kvm"
"-cpu host"
"-smp ${toString cpus}"
"-m ${toString memoryMiB}"
"-nodefaults"
"-no-user-config"
"-display none"
"-serial stdio"
"-device virtio-rng-pci"
"-drive file=${disk},if=virtio,format=qcow2,cache=writeback,discard=unmap"
# User-mode NAT: outbound works (the guest needs to reach the caches
# to rebuild itself), and only the forwarded ssh port is reachable
# from outside. Nothing the agents do in here can squat a LAN IP.
"-netdev user,id=net0,hostfwd=tcp::${toString sshPort}-:22"
"-device virtio-net-pci,netdev=net0"
"-qmp unix:${qmp},server=on,wait=off"
];
# ACPI powerdown over QMP rather than `ssh poweroff`: on a box whose
# whole job is surviving agent experiments, sshd is exactly the thing
# that may be broken when it is time to stop.
ExecStop = pkgs.writeShellScript "${unit}-stop" ''
printf '%s\n' \
'{"execute":"qmp_capabilities"}' \
'{"execute":"system_powerdown"}' \
| ${lib.getExe pkgs.socat} -t2 - UNIX-CONNECT:${qmp} || true
'';
TimeoutStopSec = "90s";
Restart = "always";
RestartSec = "5s";
StandardInput = "null";
NoNewPrivileges = true;
ProtectSystem = "strict";
ProtectHome = true;
PrivateTmp = true;
RestrictSUIDSGID = true;
};
};
# Runs on every VM start, not just after a reseed: if an agent wipes
# /etc/nix in there, the next restart puts the cache credential back.
systemd.services."${unit}-netrc" = {
description = "Copy the private cache netrc into the agent-lab VM";
wantedBy = [ "${unit}.service" ];
after = [ "${unit}.service" ];
unitConfig.ConditionPathExists = [
netrc
provisionKey
];
serviceConfig = {
Type = "oneshot";
ExecStart = pkgs.writeShellScript "${unit}-netrc" ''
set -eu
for _ in $(${lib.getExe' pkgs.coreutils "seq"} 1 60); do
if ${guestSsh} true 2>/dev/null; then
# Piped from the host's file at runtime; the credential is never
# an argument, never in a derivation, never in the store.
${guestSsh} 'umask 022; cat > /etc/nix/netrc' < ${netrc}
${guestSsh} 'systemctl restart nix-daemon' || true
echo "netrc installed in ${vm}"
exit 0
fi
${lib.getExe' pkgs.coreutils "sleep"} 5
done
echo "${vm} did not answer ssh within 5min; netrc not installed" >&2
exit 1
'';
};
};
# Reachable from lpt2 (where the hive-controller's agents live) over
# tailscale/LAN as <this host>:2222.
networking.firewall.allowedTCPPorts = [ sshPort ];
environment.systemPackages = [ reset ];
programs.ssh.extraConfig = ''
Host ${vm}
HostName 127.0.0.1
Port ${toString sshPort}
User root
StrictHostKeyChecking no
UserKnownHostsFile /dev/null
'';
};
}

View file

@ -4,6 +4,7 @@
./hardware.nix
./gitea-runner.nix
./hive-ci-runner.nix
./agent-lab-vm.nix
./k8s-test-swarm.nix
# ./vscode-server.nix
# ./hass.nix