autoUpgrade was silently failing daily on both hosts (legacy channel path,
no --flake set) without reaching the network — disabled outright since
rebuilds should only happen manually. Separately, nix-gc.timer's nightly
sweep was deleting eval-time-only flake input sources (not part of the
runtime closure), forcing the next evaluation to cold-refetch hermes-agent's
entire dependency tree from GitHub in one burst, which alone was enough to
hit the unauthenticated codeload rate limit. keep-outputs/keep-derivations
stop GC from evicting those sources.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TJS8qsYPXSUm5mdWGdaHjF
Its self-updater was independently polling GitHub for the same repo the
flake input tracks, and stacked with a runaway crash-restart loop this
exhausted the GitHub rate limit, causing unrelated flake fetches to 429.
The Nix flake input is already the source of truth for hermes's version.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F9ge2kVMD55sLJExx3v2G3
Jellyseerr was rebranded upstream to Seerr (seerr-team/seerr), and
nixpkgs now aliases services.jellyseerr to services.seerr with an
identical state directory, so this is a same-data rename. Also fixes
systemd.services.jellyseerr.after, which was a silent no-op since the
underlying unit has always been named seerr.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019MLWL7mWHCV6UGfzmhnfqN
nixos-26.05 pins immich at 2.7.5; pull 3.2.0 from pkgs.unstable
instead. The machine-learning service picks up the matching version
automatically via the package's passthru. Verified jupiter's toplevel
still builds; pgvecto.rs->vectorchord migration was already in place
so no DB migration is needed for the bump itself.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DDyhS2N9XEeMq4GisBrqyf
homepage-dashboard resolves services.yaml groups by name using first-match
lookup, so the module-contributed "Services" group and manual-services.nix's
"Services" group (two separate top-level YAML list entries) caused the
second one to be silently discarded. Merge by group name before handing the
list to services.homepage-dashboard.services.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JbLevsLozTCcCWe1SoUGe1
Add an 8G swapfile for headroom during nix builds (7.7G RAM, no prior
swap) and bring nodejs (bundles npm) into the development profile.
Also picks up latest flake.lock input updates.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01C5fRUDLRzDhoKhEqa69cyR
Hermes Agent's dashboard runs on mibook, but the homepage-dashboard
service only runs on jupiter, so wire it in as a manual cross-host
entry rather than via my.homepage.services (which is scoped per-host
and would be inert on mibook).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Sux8hmeb2qadqPqbJ1zNzL
Bind the backend to 0.0.0.0:9119 in dashboard mode so the web UI is
reachable from the local network, open the port in the firewall, and
add dashboard username/password credentials to the mibook hermes-env
secret.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Sux8hmeb2qadqPqbJ1zNzL
The fix subagent found that inputs.hermes-agent.nixosModules.default
inside `imports` causes infinite recursion in this repo's wiring
(inputs comes from _module.args, which merges after imports resolve;
self comes from specialArgs, which is available at import-time).
Tested both ways. Update docs to show self.inputs, matching commit
31e1910.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Vbvwg3Qt2u9nbkEA3wL17S
machines/configuration.nix wired hermes-agent.nixosModules.default into
mibook's mibook-only modules list. modules/environments/default.nix is
machine-agnostic, so jupiter still evaluated a
services.hermes-agent = lib.mkIf false { ... } definition for an option it
never declared as existing — NixOS errors on that regardless of mkIf's
condition, breaking nixosConfigurations.jupiter (and the repo's sole CI
job) entirely.
Fix: import the upstream module unconditionally from inside
modules/environments/hermes-agent/default.nix itself, so the option is
declared everywhere modules/environments/default.nix loads, and only ever
turned on where my.profiles.hermes-agent.enable = true (mibook). Also adds
users.users.finn.extraGroups = [ "hermes" ], required for the hermes CLI to
be usable per the corrected design. machines/configuration.nix no longer
mentions hermes-agent.
Deviates from the reconciled spec/plan text in one respect: uses `self`
(from specialArgs) instead of `inputs` (from _module.args) to reach the
flake input inside `imports`. Using `inputs` there reproduced infinite
recursion — NixOS resolves a module's imports before _module.args
assignments are merged, so an import list cannot depend on an arg that
itself comes from config. `self` is available at import-time via
specialArgs, `inputs` is not, in this repo's wiring. See report for detail.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Vbvwg3Qt2u9nbkEA3wL17S
Final review (opus) found a Critical bug (jupiter's nixosConfigurations
stopped evaluating because hermes-agent.nixosModules.default was wired
through machines/configuration.nix's mibook-only module list, while
the profile module itself is loaded on both machines) plus three
Important gaps: secrets persist to /var/lib/hermes/.hermes/.env, not
/run/secrets; Telegram/Email need explicit ALLOWED_USERS or they're
wide open / silently drop messages; addToSystemPackages's CLI needs
group membership finn didn't have. All four corrected in the docs
here; the actual code fix is a separate dispatched task.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Vbvwg3Qt2u9nbkEA3wL17S
Task 1's implementer found sops.validateSopsFiles makes secret-key
existence a build-time check (via manifest.json.drv), not an
activation-time one as the plan assumed. Corrects Step 10 to reflect
reality rather than leave future readers misled.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Vbvwg3Qt2u9nbkEA3wL17S
Task 1 is agent-executable repo wiring (flake input, profile module,
secret declaration), build-checked with nix build. Tasks 2-5 are an
operator-run runbook: OAuth login, Anthropic billing check, Telegram
bot setup, and typing real secrets into $EDITOR via sops - none of
that is meant to pass through an agent session.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Vbvwg3Qt2u9nbkEA3wL17S
Locks in host (mibook, not jupiter), deployment mode (NixOS native
module), and auth (claude setup-token / ANTHROPIC_TOKEN, Max extra
credits) after a grilling round. Builds on feat/sops-nix as-is.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Vbvwg3Qt2u9nbkEA3wL17S
Proves the sops-nix pipeline end to end on mibook: encrypts a
throwaway secret to the operator + mibook host keys, declares it in
machines/mibook/secrets.nix, and verifies it decrypts to
/run/secrets/placeholder (root:root, 0400) on nixos-rebuild switch.
Also removes the dead `let secrets = config.sops.secrets;` binding
from both environments.nix files — leftover from an earlier abandoned
plan, never referenced.
Wires sops-nix into every machine via machines/core/sops.nix so hosts
decrypt with their existing SSH host key. Adds a devShell exposing
sops/age/ssh-to-age for the operator workflow, which also fixes .envrc's
previously no-op `use flake`.
Back out the entire "choose terminal-only vs KDE at boot" work. The first
attempt (default multi-user.target + on-demand desktop) locked mibook out —
it is WiFi-only with credentials in KWallet, so a headless boot has no
network and no SSH. The follow-up specialisation redesign is no longer
wanted either.
mibook returns to plain always-KDE boot (the pre-feature baseline):
- machines/mibook/configuration.nix: drop specialisation.terminal and the
NetworkManager-wait-online override.
- machines/mibook/environments.nix: drop kde-desktop.startOnBoot.
- modules/environments/kde-desktop/default.nix: drop the startOnBoot option.
- remove the boot-desktop-choice design spec.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1wvPy5kbGp8pUQX61WBCY
finn.markwitz
