The incus-pet CLI gains three big changes that together cut per-container
disk usage by ~5GB, drop nix-rebuild + nix-copy from the deploy path, and
make `deploy` actually pull latest on every invocation. anywhen migrates
from a host-installed service to an incus-pet container in the process.
* Shared /nix/store mount. Each container gets host's /nix/store
bind-mounted read-only (with shift=true for idmapped unprivileged
containers). Zero per-container duplication of nixpkgs closures. The
CLI builds toplevels on the host; the container sees them via the
shared mount, no `nix copy` needed.
* Restart-to-deploy. switch-to-configuration is removed entirely (it
was stalling on dbus-broker reload and hitting the same lock-
contention bug across redeploys). New activation: update
/nix/var/nix/profiles/system, repoint /sbin/init through the profile
symlink, `incus restart`. The new toplevel's own init script handles
/etc setup on boot. Trade-off: ~5s of container reboot per deploy.
* Marshaling flake exposes only nixosModules.default. No nixpkgs
input, no nixosConfigurations entry. The CLI picks nixpkgs at build
time (pinned to nixos-25.11 — matches the LXC image baseline, dodges
unstable regressions) and assembles the full nixosSystem via
`nix build --impure --expr`. Lets us bump nixpkgs without touching
per-deploy state.
* `nix flake update --refresh` on every deploy. Was `nix flake lock`,
which only locks first-seen inputs — re-deploying `github:srid/anywhen`
stayed pinned at the first commit ever resolved. Now every deploy
bumps the target input to the current tip of its ref and bypasses
nix's fetcher cache.
* /sbin/init repoint. The NixOS LXC image hard-links /sbin/init to its
own toplevel's init path. Without switch-to-configuration, `incus
restart` re-execs that old init, which sets up /etc from the old
toplevel — running services keep using the OLD anywhen binary even
though the profile pointer moved. CLI now also does
`ln -sfn /nix/var/nix/profiles/system/init /sbin/init` so the profile
actually drives boot. Idempotent.
* Host gcroot via `nix-store --add-root --indirect`. Container's nix DB
doesn't know about the path (it lives behind the shared mount), so
without a host-side gcroot the host could nix-collect-garbage the
running system out from under the container. Symlink at
~/.local/state/incus-pet/<name>/system. Only the LATEST toplevel is
pinned per app — no rollback today (file a 5-line ring buffer when
rollback matters).
* anywhen flake input removed. `modules/nixos/linux/anywhen.nix` is
gone; the host config no longer imports anywhen as a flake input.
The deploy command takes the flake ref as an argument, so locking it
here was just bloat. `just pureintent anywhen-deploy` runs the
current invocation.
* pureintent ops recipes. New `configurations/nixos/pureintent/mod.just`
with anywhen-{deploy,rm,backup,restore,status,shell}. Namespaced via
`mod pureintent` so each host's recipes live behind their own prefix.
* Drop ssh from deploy path. Bootstrap still pushes the operator's
pubkey (for interactive debugging), but activation goes through
`incus exec` exclusively. No NIX_SSHOPTS, no nix-copy-via-ssh,
no host-key TOFU dance during deploys.
Tested end-to-end on pureintent: anywhen migrated from host-installed
(21-task SQLite DB at /var/lib/anywhen) to an incus-pet container with
the DB intact; hello-web deployed fresh with shared /nix/store; both
serve over the tailscale proxy at 100.122.32.106:{6111,8081}; lock
bumps + new toplevels confirmed after master advances.
* docs: Ralph report scaffolding + pureintent eval baseline (10.87s)
* cycle-1: drop nixvim — pureintent eval 10.87s → 6.98s (-36%)
nixvim's home-manager module system is by far the dominant cost of
evaluating `nixosConfigurations.pureintent`: a profile of selective
imports puts it at 3.97s of the 10.87s baseline (~36%).
Replace it with a minimal `programs.neovim.enable = true` configuration.
The runtime behaviour of `nvim` changes (no plugins, no LSP, no
mapleader/telescope/treesitter/etc.) — accepted by the user.
Measured on srid-nc (nixos, 6.12.85 kernel), 7 warm-cache runs with
`--option eval-cache false`:
baseline: 10.85 10.86 10.86 10.87 10.87 10.88 10.91 -> median 10.87s
cycle-1: 6.94 6.96 6.98 6.98 7.02 7.02 7.73 -> median 6.98s
Other configurations still evaluate:
- nixosConfigurations.naiveintent
- darwinConfigurations.infinitude-macos
- homeConfigurations."srid@zest"
Removes flake input `nixvim` (+ its `flake-parts`, `nixpkgs`, `systems`
sub-inputs from the lock).
* docs: wrap up Ralph report (10.87s -> 6.99s = -35.7%)
7 dead-ends documented so they don't have to be re-tried. Key
finding: after dropping nixvim (cycle 1), the eval floor is
home-manager's per-entry submodule materialisation
(systemd.user.services / programs.ssh.matchBlocks), not the
option-declaration count of any wrapper module. Inlining
jumphost-nix or vira moves the cost; it does not eliminate it.
- Move AI config from nix-agent-wire/srid to ./AI
- Update flake input: srid/AI -> srid/nix-agent-wire
- Update home modules to use local AI folder via flake.self.outPath
- Add opencode/web.nix module for web service configuration
- Bind to Tailscale IP (100.122.32.106) on port 4096
- Pass JUSPAY_API_KEY from agenix secret to service
- Update flake inputs: nixpkgs/unstable, nix-darwin/master, home-manager/master
