Per-run tooling verification (added in the previous commit) was hardcoded
to checking just `nix`. Turn it into a small array instead, so a future
remote build step needing another tool extends that list rather than
growing a second parallel check.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Two bugs found running scripts/create-proxmox-resource.sh against a real
Proxmox node for the first time:
- The Nix installer's single-user root path still shells out to `sudo`
to create /nix even though it already detected it's running as root,
which fails outright on a minimal, sudo-less Debian/Proxmox node.
codex-setup.sh now pre-creates /nix itself so that branch of the
installer is skipped.
- ensure_remote_repo() only ran scripts/codex-setup.sh right after a
fresh git clone, so a bootstrap that cloned the repo but then failed
installing Nix (exactly the failure above) left every subsequent run
silently building with a `nix` that was never actually installed,
since the repo already existing skipped tooling setup entirely. It
now checks `command -v nix` (via the same ensure_nix_profile used
elsewhere, since a non-interactive ssh session won't otherwise have a
single-user install on PATH) on every run and re-bootstraps if it's
missing. Both remote build heredocs also now source
scripts/lib/nix-bootstrap.sh themselves for the same PATH reason.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
New tor-relay build type (currently lxc-only) running a plain Tor
middle relay via modules/tor/enable-relay.nix, plus nyx for
interactive monitoring over the relay's control socket.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
create-proxmox-resource.sh no longer builds locally and scp's a
multi-gigabyte image over -- it now clones/pulls this repo onto the
Proxmox node itself (bootstrapping build tooling via the existing
codex-setup.sh on first use) and runs the nix build / disko image
script there, staging the result straight into the node's own import
directory. host-keys/ (gitignored) is copied over separately since a
git pull doesn't carry it. --image still uploads an explicit local
file for the case where you don't want a build at all.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
--allow-duplicate-host previously just skipped the existing-resource
check entirely, so re-running e.g. --type lxc --host docker while an
lxc-docker container already existed created a second container
sharing the same hostname/identity instead of replacing it -- both
then fight over DNS/DHCP for that hostname, and it's easy to end up
testing the stale one without realizing.
Now splits matches into "exact" (same --type as the one being
created, e.g. another lxc-docker) and "cross-type" (a different
platform sharing this host identity, e.g. a proxmox-docker VM
alongside an lxc-docker container -- a deliberate, valid coexistence
this script has never managed and still won't). Only an exact match is
destroyed and replaced, after typing the hostname back to confirm; a
cross-type match is always left untouched. Without
--allow-duplicate-host, both cases still refuse to run exactly as
before.
Verified live against pve.sweet.home: correctly split VMID 103 (a
stopped proxmox-docker VM, cross-type -- left untouched) from VMID 105
(the running lxc-docker container, exact-type -- flagged for
destroy+replace), and confirmed the destroy prompt safely aborts on a
non-matching confirmation, leaving both resources untouched.
Two compounding bugs, confirmed live on the running lxc-docker
container (vmid 102 on pve.sweet.home):
1. x-systemd.automount never works inside any Linux container --
systemd logs "Starting of <unit>.automount unsupported" for every
share and never mounts them. modules/docker/mount-data.nix and
modules/raspi/mount-data.nix now key off config.boot.isContainer
(set true by nixpkgs' proxmox-lxc.nix) to mount eagerly with
`nofail` there instead, while VM-based docker targets keep automount
unchanged.
2. The container's Proxmox `features` never included `mount=nfs`, so
AppArmor blanket-denies the nfs/rpc_pipefs mount syscalls NFS
needs ("permission denied"). scripts/env.sh's
PROXMOX_DEFAULT_LXC_FEATURES now includes mount=nfs;nfs4 for future
lxc-* containers -- the semicolon required quoting the --features
value in create-proxmox-resource.sh's remote pct-create command,
since it's sent as a raw string for the remote shell to parse and an
unquoted `;` would be read as a command separator.
The already-running container needs a matching `pct set --features`
plus a restart to pick this up -- that's an operator step outside this
repo.
Companion to rotate-admin-key.sh: copies whatever age identity sops/age
itself would resolve (or an explicit --key-file) to a given destination
path with 0600 perms, validating it's a real identity and round-tripping
the derived public key before/after the write so a corrupted copy is
caught immediately rather than discovered later during a restore.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Automates the manual steps sync-host-keys.sh/create-proxmox-resource.sh
print when they bootstrap a fresh, not-yet-trusted age key: verifies a
backed-up key matches the current &admin entry, swaps in a new key, and
re-encrypts every secrets/*.yaml. Explicitly cds into repo_root before any
sops call, since sops resolves .sops.yaml by walking up from cwd rather
than from the target file's path -- confirmed via a scratch-repo test that
running from elsewhere would otherwise silently rotate against the wrong
config.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Same class of problem as the deployedTargets/README fixes: hand-maintained
prose that drifted from reality and nobody was obligated to update.
- CLAUDE.md: "18 hosts" was a stale hardcoded count (actually 20); reworded
to not need updating as hosts are added. Also added the missing
tailscale-exit-node build type to a list that had it everywhere else in
the file except one bullet.
- AGENTS.md: same missing tailscale-exit-node build type.
- docs/auto-installer.md: the hand-enumerated lxc-* list was missing
lxc-tailscale-exit-node.
- flake-target-refactor-spec.md: added a "Status: implemented" note so this
completed historical spec (referenced elsewhere purely for rationale)
can't be mistaken for an open plan with unresolved Open Questions.
- remove-sensetive-info-refactor.md: the "Definition of done" checklist was
entirely unchecked despite most of the work being done. Checked off what's
actually done (sops-nix migration, history scrub just performed, the
pre-commit gitleaks hook), and left rotation of the GitHub PAT found in
history explicitly flagged as the one still-open item -- an operator
action against GitHub, not something this repo can attest to itself.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The remote bash script run over SSH ended with a for-loop whose last
statement was `[[ "$n" == "$target" ]] && echo ...`. When the last
VM/CT checked on the node didn't match --host, that test evaluated
false and became the exit status of the whole remote script (1) --
which the wrapper then misreported as "couldn't reach the node",
even though SSH connectivity and the check itself were both fine.
The actual signal is the script's stdout, not its exit code, so end
it with an explicit exit 0.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Three chunks of copy-pasted logic were drifting across scripts/*.sh:
- codex-setup.sh and codex-maintenance.sh each carried an identical
NIX_CONFIG bootstrap + ensure_nix_profile() -> scripts/lib/nix-bootstrap.sh
- sync-host-keys.sh and prepare-host-key.sh each ran the same
ssh-keygen/ssh-to-age nix-shell invocations -> scripts/lib/ssh-host-keys.sh
(prepare-host-key.sh now also calls env.sh's nix_extra_opts before using
them, closing a gap where it alone skipped the nix-cache reachability
check env.sh exists for)
- the "list nixosConfigurations attrNames" / "get one target's hostName"
nix eval pattern was repeated across codex-setup.sh, codex-maintenance.sh,
sync-host-keys.sh and create-proxmox-resource.sh (the latter twice, in
its own --list and --host lookup) -> scripts/lib/nix-eval.sh, which also
centralizes the --no-use-registries --no-accept-flake-config flag pair
used on every such call
Verified against the real flake/node config (nix is available here):
create-proxmox-resource.sh --list for both --type lxc/vm, a full
--dry-run create, and prepare-host-key.sh generating and cleaning up a
real key/age-pubkey pair.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Same problem as the deployedTargets removal, just in markdown instead of
Nix: which variant of a buildtype is actually deployed is live
infrastructure state, and a committed table can't stay accurate as that
changes -- it already required a manual edit on every migration and had
drifted before. Keep only what doesn't rot: what each target is for, and
stable naming history. Point at the live node / /etc/flake-target instead
for actual deployment status.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
variables.nix's deployedTargets was a manually-maintained list with no
enforcement keeping it in sync with reality -- it caused two separate
false refusals in a row (naming a VM as deployed well after it had been
destroyed, then matching a target against itself once the list was
"corrected"). Static files can't track whether a resource still actually
exists.
create-proxmox-resource.sh's duplicate-host guard now queries the
Proxmox node directly (qm/pct's own name/hostname config, matched
against --host) instead. Also fixes a gap in that live check: it
originally swallowed ssh failures and would have silently treated "can't
reach the node" the same as "checked, nothing there" -- it now refuses
instead of guessing when the node can't be reached.
deployedTargets is removed entirely from variables.nix since nothing
else in the repo consumed it once this script no longer does; README.md's
Hosts table remains the sole source of truth for "(real, deployed)"
status. CLAUDE.md and the script's own --help/comments updated to match.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The duplicate-host check in create-proxmox-resource.sh compared by
hostName only, so it fired even when the target being created was
exactly the one variables.nix's deployedTargets already names (e.g.
rebuilding lxc-nix-cache after destroying its old container to pick up
new sops secrets) -- there's no other machine at risk of an identity
collision in that case, just the normal redeploy workflow. Skip the
check when dt == flake_target; the later VMID-existence check still
guards against clobbering a resource that's actually live on the node.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The old proxmox-nix-cache VM was destroyed and nix-cache is being
redeployed as an LXC container going forward. Without this update,
create-proxmox-resource.sh's duplicate-host check (which only reads this
static list, not live Proxmox state) kept refusing to create
lxc-nix-cache even though nothing named nix-cache actually exists on the
node anymore.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Distributed builds failed with "Host key verification failed" on any
client that had never manually SSH'd to nix-cache before, since
nothing populated root's known_hosts for it. Wire nix-cache's host
public key into programs.ssh.knownHosts via a new vars.nixCacheHostKey
so every client picks it up automatically on rebuild.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
lxc-* hosts need NIXOS_HOST_KEYS_DIR + --impure to bake in a pre-seeded
SSH host key, otherwise sops-nix's .sops.yaml recipient never matches
and every secret permanently fails to decrypt on first boot. That
invocation is easy to forget, so wrap it as `buildImage <flake-target>`
alongside the existing Switch-nix/Test-nix helpers.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The reachability check used `cat < /dev/tcp/${NIX_CACHE_HOST}/22`, which
blocks forever reading for EOF that never comes -- sshd sends its banner
and then holds the connection open waiting for the client to speak next.
Every single check hit the 3s timeout and reported "unreachable"
unconditionally, regardless of whether the remote builder was actually up.
Confirmed live: a plain TCP connect (`exec 3<>/dev/tcp/...`, no read)
returns in ~60ms against a healthy nix-cache instead of always timing out.
Fixing that exposed a second, previously-dormant bug: `printf -v
NIX_EXTRA_OPTS '%q ' "${NIX_OPTS[@]}"` on a genuinely empty NIX_OPTS array
still runs one format pass and yields the literal `'' ` rather than an
empty string. A subprocess (e.g. sync-host-keys.sh) reusing this
process's decision via `eval "NIX_OPTS=(${NIX_EXTRA_OPTS})"` then rebuilt
a 1-element array holding an empty string instead of a 0-element array,
which broke `nix-shell "${NIX_OPTS[@]}" -p <pkg>` with a bogus positional
argument the moment NIX_OPTS was legitimately empty (nix-cache reachable)
-- something the first bug had made impossible to ever hit before.
Also adds a couple of retries (1s apart) to both checks as a secondary
safety net against genuine multi-second blips, on top of fixing the
checks themselves.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The tarball/disko-image build previews were hardcoded strings that never
included ${NIX_OPTS[@]}, so --dry-run always showed the same "would build"
command whether nix-cache's substituter/remote-builder got disabled by
nix_extra_opts's reachability probe or not -- the actual (non-dry-run)
build commands already applied it correctly, only the preview lied.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
services.beszel.agent runs under DynamicUser=true with ProtectSystem =
"strict" and no StateDirectory, so /var/lib/beszel-agent -- where the
agent persists the fingerprint that locks its hub pairing to this
machine (github.com/henrygd/beszel/discussions/1542) -- was never
actually writable. Every restart silently failed to persist it and
regenerated a fresh one in memory, permanently desyncing from whatever
the hub had on record after the very first successful pairing. Affects
every host importing modules/beszel/enable-agent.nix (nix-cache, server),
not just full container rebuilds.
Found via nix-cache showing "fingerprint mismatch" after being rebuilt
post-outage; confirmed server was silently exposed to the same bug, just
hadn't restarted since its first pairing. Fixed by declaring
StateDirectory so systemd gives the dynamic user real persistent storage.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
--name (used as pct/qm create's --hostname/--name) defaulted to
$flake_target (e.g. "lxc-nix-cache"), not $host (e.g. "nix-cache"). Since
proxmoxLXC.manageHostName pulls the guest's real networking.hostName
straight from Proxmox's own container config, this silently overrode
host.nix's hostName with a build-type-specific name. Default --name to
--host instead, so the guest's identity matches host.nix regardless of
which platform variant built it.
Found by spinning up a fresh lxc-nix-cache test container and noticing its
hostname was "lxc-nix-cache" instead of "nix-cache".
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
New build type dedicated to Tailscale exit-node capability, wired up for
linode/proxmox/lxc like every other build type (the lxc variant is the one
actually intended for deployment). Kept separate from the "server" host
rather than bundling exit-node capability onto it.
Trimmed modules/tailscale/exit-node.nix down to pure exit-node behavior:
dropped the old --advertise-routes=${vars.lanCidr} bundling (meaningless
for a Linode-hosted VPS with no path to the LAN), and switched
extraUpFlags -> extraSetFlags. Confirmed against nixpkgs' tailscale.nix
that extraUpFlags is only applied by tailscaled-autoconnect, which itself
only runs when services.tailscale.authKeyFile is set -- nothing in this
repo sets one, so the old flags would never have actually been applied.
extraSetFlags runs unconditionally via tailscaled-set on every boot, so
--advertise-exit-node self-reapplies once the operator has done the
one-time manual `tailscale up` auth.
Verified: all three new targets eval cleanly, nixpkgs-fmt/statix clean,
and a dry-run build of lxc-tailscale-exit-node's tarball resolves its full
closure including tailscaled-set.service.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01La55Nsss8jZ7ZuzUV9mfot
virtualisation/proxmox-lxc.nix registers the Nix store DB via a systemd
service, never an activation script -- so neededForUsers sops secrets
(password hashes) and the user-creation step that consumes them never ran
on a real first boot, leaving /etc/shadow stuck with build-time placeholder
entries. boot.postBootCommands looked like the right hook (stage-2-init.sh
does invoke it) but switch-to-configuration behaves unreliably that early,
before systemd itself is up. Fixed with a genuine oneshot systemd service,
gated by ConditionPathExists so it only ever runs once.
Confirmed live via a from-scratch destroy+rebuild+redeploy of the
lxc-nix-cache test container: real password hashes applied automatically,
systemctl is-system-running -> running, zero failed units.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01La55Nsss8jZ7ZuzUV9mfot
The real root cause behind the original nix-cache 502, traced all the way
through: modules/installer/host-keys.nix (which NIXOS_HOST_KEYS_DIR=...
--impure actually wires up) is only ever imported by the installer's own
modules/installer/common.nix -- modules/platforms/lxc.nix, which every
real lxc-* host build actually uses, never imported anything like it.
docs/auto-installer.md previously claimed NIXOS_HOST_KEYS_DIR bakes a key
into lxc-* tarballs "the same way it does for the ISO/PXE installer
images" -- that was never actually true; I wrote it without verifying the
mechanism existed for lxc.nix specifically.
In practice this meant every lxc-* container booted with a freshly
self-generated SSH host key that could never match whatever .sops.yaml
actually trusts for that target, so *every* secret -- not just
cache-priv-key -- silently failed to decrypt. No error surfaces in the
boot log for this: the activation step that installs secrets only runs
on a genuinely fresh first activation and silently no-ops once
/run/current-system already exists, so by the time anyone looks the
window has closed. Found by manually invoking sops-install-secrets
directly: "Error getting data key: 0 successful groups required, got 0".
Fixed by giving modules/platforms/lxc.nix the same key-baking mechanism
the installer has, but keyed to its own exact flake target and placing
the key directly at /etc/ssh/ssh_host_ed25519_key (no copy step to stage
for, unlike the installer's /etc/host-keys/ staging area -- an lxc-*
tarball has no install step). The target name comes in via
specialArgs.flakeTarget (new, set by flake.nix's mkTarget) rather than
being read back from config.environment.etc."flake-target" -- reading
that back from within a module that also contributes to
environment.etc is circular (confirmed: "infinite recursion
encountered").
Verified live end-to-end against the real test container (lxc-nix-cache,
VMID 100 on pve.sweet.home): destroyed it, rebuilt the tarball fresh with
the fix, recreated it, and confirmed /run/secrets/ now has all three
secrets this host needs (beszel-token, cache-priv-key, nix-github-token),
nix-serve is active (running), and curl http://localhost/nix-cache-info
succeeds both directly and through nginx.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01La55Nsss8jZ7ZuzUV9mfot
exit-node.nix advertised --advertise-exit-node/--advertise-routes but
never enabled IP forwarding, so it wouldn't actually route traffic.
services.tailscale.useRoutingFeatures = "server" is the built-in way to
get the sysctls right without duplicating modules/networking/enable-ip-forwarding.nix.
Also opens the firewall for direct (non-DERP-relayed) peer connections,
and now imports enable-service.nix instead of duplicating `enable = true`,
so "plain tailscale" and "tailscale exit node" are two composable modules.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Found and fixed live against a real test container (VMID 100, lxc-nix-cache
on pve.sweet.home) after the previous pct-restore-to-pct-create fix still
produced a container that booted into garbled console output:
1. pct create's own CLI default for --unprivileged is privileged (unlike
the web UI, whose checkbox defaults the other way), but
modules/platforms/lxc.nix sets proxmoxLXC.privileged = false, so the
image assumes it's running unprivileged. Real mismatch -- now passes
--unprivileged 1 explicitly.
2. The actual root cause of the garbled console: modern (v247+) systemd
routinely uses nested user namespaces and credential mounts (even
plain getty units, via LoadCredential=-style mechanisms), which
AppArmor's default LXC confinement denies without --features
nesting=1,keyctl=1. Confirmed via the host's kernel audit log: every
getty unit was crash-looping on a denied /run/credentials/* mount
every ~3s, and core services like nsncd failed userns_create the same
way -- the system never finished activating. Fixed live (pct set +
restart on the running test container) before committing the script
change: systemctl is-system-running went from never completing to
"running" with zero failed units.
3. --memory doesn't touch swap -- confirmed live it silently stayed at
Proxmox's own 512M default with --memory 2048. Now defaults --swap to
whatever --memory resolves to.
docs/auto-installer.md's manual pct create walkthrough gets the same
fixes, with the "why" for each flag, since a human following it by hand
would hit the identical bugs.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01La55Nsss8jZ7ZuzUV9mfot
Two independent problems found while actually running the script:
1. nix build/nix-shell retry each unreachable substituter/builder up to
5x with backoff, per store path -- with nix-cache down this compounds
into minutes of noise. scripts/env.sh gains nix_extra_opts(), which
probes http://nix-cache and nixremote@nix-cache:22 once via plain
curl/TCP (bypassing Nix's own retry logic entirely -- confirmed
nix store ping still retries 5x even with a short connect-timeout)
and exports the decision so create-proxmox-resource.sh and the
sync-host-keys.sh subprocess it shells out to both reuse it instead
of probing independently.
2. The actual failure: "archive contains no configuration file". pct
restore expects a vzdump backup archive with embedded config;
config.system.build.tarball is a plain CT template tarball -- wrong
Proxmox mechanism entirely. Fixed to pct create against it as a vztmpl
template instead, uploaded to /var/lib/vz/template/cache/ rather than
/var/lib/vz/dump/. This same wrong claim had propagated into
docs/auto-installer.md, README.md, and CLAUDE.md from when the script
was first written -- corrected everywhere.
Also: checks for an already-uploaded image on the node (fixed
<flake_target>.tar.xz/.raw naming) before building, skipping build+upload
entirely if found (--force-rebuild to always rebuild).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01La55Nsss8jZ7ZuzUV9mfot
sync-host-keys.sh: generates/registers SSH host keys and their
.sops.yaml/secrets/*.yaml recipients for flake targets, idempotently.
--all, <target>, --remove, --regenerate-all-keys, all with --dry-run
(verified zero-side-effect via a sandboxed git-status check across every
mode). Only ever touches anchors with a corresponding host-keys/ file --
&admin and any hand-registered real-host anchor are never listed,
removed, or regenerated. Supersedes running prepare-host-key.sh one host
at a time for any target that already has a flake entry.
create-proxmox-resource.sh: builds a lxc-*/proxmox-* target's
tarball/disk image and creates it on a real Proxmox node, or reconfigures
an existing resource's cores/memory/disk (--modify, always requires
typing the VMID back to confirm). Refuses to create a new resource for a
VMID that already exists, and refuses to duplicate a host identity that
already has a real deployment elsewhere (variables.nix's new
deployedTargets, checked by hostName so it also catches cross-platform
duplicates) unless --allow-duplicate-host is passed. --dry-run throughout.
scripts/env.sh centralizes the Proxmox connection config both scripts
(and future ones) share. Also fixes an unrelated gap found along the way:
proxmox-* Disko image builds write their .raw file straight into the
repo root, and .gitignore never covered it.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01La55Nsss8jZ7ZuzUV9mfot
nix-serve's secretKeyFile was a manual, undocumented-outside-a-comment
`nix-store --generate-binary-cache-key` step per host -- easy to miss on
a fresh nix-cache instance (as lxc-nix-cache testing just found: systemd
fails the unit with EXIT_CREDENTIALS when LoadCredential can't find the
source file, which nginx then reports as a 502 from clients). It also
can't be regenerated per-host safely: modules/nix-cache/client.nix
hardcodes every client's trust in one specific public key, so every
nix-cache instance has to share the exact same keypair.
Sourced from secrets/nix-cache.yaml's new cache-priv-key entry instead,
via the same sops-nix pattern every other secret in this repo already
uses. Verified the added value derives to the exact public key
modules/nix-cache/client.nix already trusts before committing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01La55Nsss8jZ7ZuzUV9mfot
beatzaplentyandClaude Sonnet 5