ci: Add docs assignment-table auto-updater

Regenerate the marked network-assignment tables in docs/ from nixos.allAssignments (nix run .#update-docs-assignments, registered in pkgs/default.nix), preserving hand-written Notes cells. The workflow runs on pushes to the docs branch and commits any table updates back; flip its trigger to master when the docs branch merges.
docs: Document the deployment
2026-07-26 19:16:43 +01:00 · 2026-07-26 19:16:43 +01:00 · 2026-07-26 18:55:20 +01:00 · 2026-07-23 21:35:50 +01:00 · 2026-07-23 21:29:16 +01:00 · 2026-07-23 01:37:33 +01:00
57 changed files with 3519 additions and 65 deletions
@@ -0,0 +1,38 @@
+name: Update docs assignments
+
+on:
+  push:
+    branches: [docs-kimi-new]
+
+jobs:
+  update:
+    if: "!contains(github.event.head_commit.message, 'docs: update assignment tables')"
+    runs-on: ubuntu-26.04
+    permissions:
+      contents: write
+    steps:
+      - uses: actions/checkout@v6
+
+      - uses: cachix/install-nix-action@v31
+        with:
+          github_access_token: ${{ secrets.GH_PULL_TOKEN }}
+          extra_nix_config: |
+            extra-substituters = https://nix-cache.nul.ie
+            extra-trusted-public-keys = nix-cache.nul.ie-1:BzH5yMfF4HbzY1C977XzOxoPhEc9Zbu39ftPkUbH+m4=
+
+      - name: Update assignment tables
+        run: nix run .#update-docs-assignments
+
+      - name: Commit and push if changed
+        env:
+          GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
+          REPO_URL: ${{ gitea.repositoryUrl }}
+        run: |
+          git config user.name "github-actions[bot]"
+          git config user.email "github-actions[bot]@users.noreply.github.com"
+          git remote set-url origin "${REPO_URL/https:\/\//https:\/\/oauth2:${GITEA_TOKEN}@}"
+          git add docs/
+          if ! git diff --cached --quiet; then
+            git commit -m "docs: update assignment tables"
+            git push
+          fi
@@ -2,6 +2,13 @@

 This file provides guidance to coding agents when working with code in this repository.

+`CLAUDE.md` at the repo root is a symlink to this file — edit `AGENTS.md`, not the symlink (some
+tools refuse to write through a symlink and will error on `CLAUDE.md`).
+
+**Prefer this file over agent memory.** When you learn something durable about this repo — a
+convention, a workflow gotcha, a design rationale — record it here (or in a repo doc this file points
+to, e.g. `docs/sites/home/switches.md`), not in agent memory. AGENTS.md is versioned and shared; memory is not.
+
 ## Overview

 Personal Nix flake managing NixOS systems and home-manager configurations for a set of
@@ -28,21 +35,33 @@ Common ones:
  Pass the flake-qualified node, e.g. `deploy .#git`. The deploy node name is **always** the system
  name (`deploy-rs.nix` keys nodes directly off `nixos.systems` / `home-manager.homes`); a system is
  only a deploy target when `config.my.deploy.enable` is true (defaults true; auto-disabled for dev
-  VMs and containers).
+  VMs and containers). Pass `--boot` to stage a config as the boot default **without** live-switching
+  (`deploy --boot .#<host>`) — the box keeps running its current generation until it reboots. Use this
+  when a live `switch` would break connectivity mid-change (e.g. a router's WAN VLAN rework), then
+  reboot to cut over.
 - `ssh-machine <name> [cmd]` — SSH to a NixOS system or home-manager config by name. Resolves the
  target and ssh options (identity, port) from its deploy-rs node, so it needs `my.deploy.enable`
  (same gate as `deploy`). Boxes default to the `fish` login shell, so pipe multi-statement remote
  scripts through `bash` (e.g. `ssh-machine <name> bash -s < script.sh`) rather than `&&`/`for`.
+  If outbound SSH hangs at the publickey step (flaky `ssh-agent`), disable the agent for the call:
+  `SSH_AUTH_SOCK= ssh-machine …` (or add `-o IdentityAgent=none` to a raw `ssh`).
 - `ragenix` — edit age secrets using `.keys/dev.key` as identity (see Secrets).
 - `repl` — `nix repl .#`.
 - `update-nixpkgs` / `update-home-manager` — bump pinned inputs.

 Check everything (what CI runs): `nix flake check --no-build`.
 CI builds each attr of `.#ci.x86_64-linux` (systems, homes, packages, shell) and pushes to the
-Harmonia binary cache; see `.gitea/workflows/ci.yaml` and `ci/push-to-cache.sh`.
+Harmonia binary cache; see `.gitea/workflows/ci.yaml` and `ci/push-to-cache.sh`. A separate
+workflow (`.gitea/workflows/update-docs.yaml`) regenerates the network-assignment tables in `docs/`
+via `nix run .#update-docs-assignments`.

 ## Architecture

+The mechanics in this section have expanded human-readable write-ups under `docs/`:
+`docs/architecture.md` (module system), `docs/networking.md` (assignments, topology, meshes) and
+`docs/deployment.md` (deploy-rs, devshell, secrets, CI). This section stays the terse agent
+version; consult those for depth.
+
 ### The custom module system
 `flake.nix` does **not** call `nixosSystem` per host directly. Instead it `evalModules` over
 `./nixos`, `./home-manager`, `./deploy-rs.nix`, and the per-host files listed in the `configs`
@@ -92,6 +111,12 @@ Per-host configs live under `nixos/boxes/<host>` (some are single `.nix` files,
 with nested VMs/containers under e.g. `colony/vms`). Many "systems" are VMs or containers managed
 via the `vms` / `containers` modules and the `l2mesh` VXLAN module.

+For a human-readable map of what is actually deployed (per-box roles, services and networking),
+see `README.md` and `docs/` (index at `docs/README.md`; box pages under `docs/sites/`,
+`docs/remote/`, `docs/mobile/`). Keep these in sync when adding, removing or repurposing a box or
+service. The network-assignment tables in box pages are CI-generated from `allAssignments` (the
+`<!-- assignments: <name> -->` markers) — write the prose and let the updater refresh the tables.
+
 ### Home routers (`nixos/boxes/home/routing-common`)
 The two home routers, `river` and `stream`, share `routing-common`, which is a **function of an
 `index`** (`import ../../routing-common 0` for river, `1` for stream). The index derives per-box
@@ -115,6 +140,20 @@ box file, not `routing-common`.
  easy to mix up. Set an interface MTU via the `.network`'s `linkConfig.MTUBytes` (`[Link]`), not
  `netdevConfig` (`[NetDev]` rejects `MTUBytes`).

+### Home switches (`jim` / `dave` / `brian`)
+The home boxes and the Digiweb WAN hang off hand-configured switches that are **not** managed by
+this flake: `jim` and `dave` (MikroTik, RouterOS) and `brian` (Ubiquiti, UniFi). The full topology,
+VLAN map, and the ONT/WAN path live in **`docs/sites/home/switches.md`** — read it before
+touching anything WAN/VLAN-related, and update it when the switch layout changes.
+- **Access:** the switches resolve by **short hostname** on the home network (the routers serve
+  their records in the home zone — `routing-common/dns.nix`: `jim`/`dave`/`brian`). From a home box,
+  SSH to the MikroTiks as `admin`/`admin` (e.g. `ssh admin@jim`); `brian` is configured via the
+  UniFi controller, not a CLI.
+- **Changing switch config is out-of-band and hard to revert — always confirm before applying:**
+  print the affected menu, make the change, then re-verify. The nix config and the switches must
+  agree on VLAN numbering (e.g. `lib.my.c.home.vlans`), so a switch-side change usually pairs with a
+  box change; `docs/sites/home/switches.md` documents the switch layout and per-switch config for the WAN design.
+
 ## Secrets

 age-encrypted secrets in `secrets/`, managed with **ragenix**. Each module declares
@@ -130,6 +169,9 @@ private keys) is required for editing secrets, deploying, and running dev VMs.
  **Ask before running `fmt`** — some files aren't canonically formatted, so `fmt` can reindent a
  whole file and bury a logical change in whitespace churn. Match the surrounding style by hand and
  leave formatting to the user unless they ask.
+- Comment where it genuinely aids understanding, but not for trivial/obvious code — match the file's
+  existing (fairly sparse) comment density. When adding something general, comment its general
+  purpose, not the specific change or one-off reason it was introduced for.
 - Prefer `lib.my` helpers (`mkOpt'`, `mkBoolOpt'`, `mkDefault'`) and `lib.my.c` constants over
  reimplementing.
 - New shared functionality → a module in `*/modules/` + entry in `_list.nix`, options under `my.*`.
@@ -0,0 +1 @@
+AGENTS.md
@@ -0,0 +1,89 @@
+# nixfiles
+
+Personal Nix flake managing every machine I run: hosted servers, home
+infrastructure, routers, a remote site, VPSes and personal workstations. It is
+built around a **custom module system** layered on top of NixOS and
+home-manager rather than the stock per-host `nixosConfigurations` pattern.
+
+For day-to-day commands and a deeper explanation of the module system,
+conventions and secrets, see [`AGENTS.md`](AGENTS.md). This README is the map of
+**what is actually deployed**; the per-machine details live under [`docs/`](docs)
+(start at [`docs/README.md`](docs/README.md)).
+
+> **Note:** This documentation (the README and everything under `docs/`) is a
+> work in progress and was **agent-generated** from the repository. It may be
+> incomplete or out of date — treat the Nix configuration as the source of truth.
+
+## The boxes at a glance
+
+Machines are grouped by deployment/location. Each group has its own directory
+under `docs/` with a `README.md` overview and one file per machine.
+
+| Group | What it is | Docs |
+| --- | --- | --- |
+| **colony** | Hosted dedicated server in Amsterdam (`ams1`). A VM host running the public-facing infrastructure: routing, web, git, media, object storage, chat, game servers. | [`docs/sites/colony/`](docs/sites/colony) |
+| **home** | Home network: a VM host (`palace`), the home routers, storage, Home Assistant, and personal desktops — plus the hand-configured switches tying it together. | [`docs/sites/home/`](docs/sites/home) |
+| **remote** | Edge VPSes (`britway`, `britnet`) and the remote `kelder` site. | [`docs/remote/`](docs/remote) |
+| **mobile** | The `tower` laptop. | [`docs/mobile/`](docs/mobile) |
+
+The custom installer image is documented at [`docs/misc/installer.md`](docs/misc/installer.md).
+Cross-cutting topics — the module system, the network, deployment — live next to
+the index: [`docs/architecture.md`](docs/architecture.md),
+[`docs/networking.md`](docs/networking.md), [`docs/deployment.md`](docs/deployment.md).
+
+## The "big machine" pattern
+
+The larger sites (`colony`, `home`) all follow the same shape:
+
+```
+physical host (VM host)
+├── VM ── thing impractical to containerise (router, storage, podman host, …)
+├── VM ── container host ──┬── NixOS container ── application
+│                          ├── NixOS container ──┬── application
+│                          │                     ├── application
+│                          │                     └── application
+│                          └── …
+└── VM ── …
+```
+
+- A **physical host** (`colony`, `palace`) does little itself beyond running
+  **VMs** via the custom `my.vms` module (QEMU + systemd units, LVM-backed
+  disks).
+- **VMs** exist for things that are impractical to put in a container — kernel
+  features, separate networking, podman/OCI workloads, foreign OSes.
+- One VM is usually a **container host** (`shill` on colony, `sfh` on home; and
+  `kelder` directly). It runs **NixOS containers** via the custom `my.containers`
+  module (systemd-nspawn based, each with its own address on a bridge).
+- **Most applications live in those NixOS containers.** A container isn't limited
+  to a single application — it commonly hosts a **group of related applications**
+  that belong together (e.g. `jackflix` runs Jellyfin, the *arr stack,
+  Transmission, PhotoPrism and copyparty; `object` runs MinIO, Harmonia, HedgeDoc
+  and wastebin). Each container is a first-class entry in the docs.
+
+Networking between everything is largely defined by per-system `assignments`
+(IPs/prefixes) plus an L2 VXLAN mesh (`my.vpns.l2`, AS211024) that ties the edge
+routers together. See [`docs/networking.md`](docs/networking.md) for the full
+picture and [`AGENTS.md`](AGENTS.md) for the mechanics.
+
+## Repo layout
+
+```
+README.md           <- you are here
+nixos/
+  boxes/            per-machine configuration ("boxes")
+    colony/         colony host + its VMs (vms/) + shill's containers
+    home/           palace host + its VMs, routing-common, plus stream, castle
+    britway/        britnet.nix, kelder/, tower/, installer.nix …
+  modules/          shared NixOS modules (my.* options); registered in _list.nix
+home-manager/       home-manager modules + configs
+lib/                lib.my helpers, constants (lib.my.c), net/dns helpers
+pkgs/               custom packages (overlays.default)
+secrets/            age-encrypted secrets (ragenix)
+devshell/           devshell commands (build/deploy/check/ssh helpers)
+ci/                 CI helpers (binary-cache push, docs assignment-table updater)
+docs/               deployment documentation (index at docs/README.md)
+```
+
+A machine is wired into the flake by adding its box file to the `configs` list
+in `flake.nix`. See [`AGENTS.md`](AGENTS.md#architecture) for how `evalModules`
+turns these into `nixosConfigurations`, `homeConfigurations` and `deploy` nodes.
@@ -0,0 +1,128 @@
+#!/usr/bin/env python3
+"""Update assignment tables in docs/ from nixos.allAssignments."""
+
+import json
+import re
+import subprocess
+import sys
+from pathlib import Path
+
+DOCS_DIR = Path("docs")
+HEADER = "| Name | Assignment | IPv4 | IPv6 | Domain | Notes |"
+
+
+def fmt_ip(ip: dict) -> str:
+    addr = ip.get("address")
+    if addr is None:
+        return "—"
+    mask = ip.get("mask")
+    gateway = ip.get("gateway")
+    parts = [f"{addr}/{mask}"]
+    if gateway is not None:
+        parts.append(f"gw {gateway}")
+    return f"`{' '.join(parts)}`"
+
+
+def parse_notes(lines: list[str]) -> dict[str, str]:
+    """Extract existing Notes indexed by Assignment from a marked table."""
+    notes: dict[str, str] = {}
+    for line in lines:
+        stripped = line.strip()
+        if not stripped.startswith("|"):
+            continue
+        cells = [c.strip() for c in stripped.split("|")]
+        # Splitting "| a | b |" produces ['', 'a', 'b', '']
+        if len(cells) < 8:
+            continue
+        key = cells[2]
+        if key == "Assignment":
+            continue
+        notes[key] = cells[6]
+    return notes
+
+
+def render_table(assignments: dict, notes: dict[str, str]) -> list[str]:
+    lines = [
+        HEADER,
+        "|---|---|---|---|---|---|",
+    ]
+    for key, a in assignments.items():
+        name = a.get("name", key)
+        alt = ", ".join(a.get("altNames", []))
+        if alt:
+            name = f"{name} ({alt})"
+        lines.append(
+            "| "
+            + " | ".join(
+                [
+                    name,
+                    key,
+                    fmt_ip(a.get("ipv4", {})),
+                    fmt_ip(a.get("ipv6", {})),
+                    a.get("domain") or "—",
+                    notes.get(key, ""),
+                ]
+            )
+            + " |"
+        )
+    return lines
+
+
+def process_file(path: Path, all_assignments: dict) -> bool:
+    box_name = path.stem
+    if box_name not in all_assignments:
+        return False
+
+    text = path.read_text()
+    lines = text.splitlines()
+
+    marker_re = re.compile(r"^<!--\s*assignments:\s*(\S+)\s*-->$")
+    for i, line in enumerate(lines):
+        m = marker_re.match(line.strip())
+        if m and m.group(1) == box_name:
+            end_idx = None
+            for j in range(i + 1, len(lines)):
+                if lines[j].strip() == "<!-- assignments-end -->":
+                    end_idx = j
+                    break
+            if end_idx is None:
+                return False
+
+            old_inner = lines[i + 1 : end_idx]
+            notes = parse_notes(old_inner)
+            new_table = render_table(all_assignments[box_name], notes)
+            new_lines = (
+                lines[: i + 1]
+                + ["<!-- assignments-start -->"]
+                + new_table
+                + ["<!-- assignments-end -->"]
+                + lines[end_idx + 1 :]
+            )
+            if new_lines != lines:
+                path.write_text("\n".join(new_lines) + "\n")
+                return True
+            return False
+
+    return False
+
+
+def main() -> int:
+    result = subprocess.run(
+        ["nix", "eval", ".#nixfiles.config.nixos.allAssignments", "--json"],
+        capture_output=True,
+        text=True,
+        check=True,
+    )
+    all_assignments = json.loads(result.stdout)
+
+    changed = False
+    for path in sorted(DOCS_DIR.rglob("*.md")):
+        if process_file(path, all_assignments):
+            print(f"updated {path}")
+            changed = True
+
+    return 1 if changed else 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
@@ -0,0 +1,96 @@
+# Deployment documentation
+
+> **Note:** these pages are a work in progress and were **agent-generated** from the repository.
+> They may be incomplete or out of date — treat the Nix configuration as the source of truth.
+
+This directory documents the boxes managed by this flake: their roles, network assignments,
+hierarchy, and the services they run. For the mechanics of the repo itself (conventions, module
+system internals for contributors, agent guidance), see [`AGENTS.md`](../AGENTS.md).
+
+The two big sites follow the pattern:
+
+```
+physical host (VM host)
+└── VM (for things impractical to containerise)
+    └── container host VM
+        └── NixOS containers (one per application group)
+```
+
+Not every box fits this pattern, but **colony** and **home** are organised this way.
+
+## Cross-cutting topics
+
+- [`architecture.md`](architecture.md) — the custom module system, `my.*` namespace, multiple
+  nixpkgs channels, shared module inventory.
+- [`networking.md`](networking.md) — network assignments, domains, site topologies, router HA,
+  the AS211024 L2 mesh, BGP, WireGuard, Tailscale.
+- [`deployment.md`](deployment.md) — deploy-rs, devshell commands, secrets workflow, CI.
+
+## Site: colony (Amsterdam)
+
+Physical host and public-infrastructure hub — see [`sites/colony/README.md`](sites/colony/README.md).
+
+```
+colony (physical VM host, ams1)
+├── estuary ── edge router: WAN, firewall/NAT, DNS, BGP (AS211024), WireGuard
+├── shill ──── NixOS container host ──┬── middleman    (reverse proxy, ACME, nginx-sso, librespeed)
+│                                     ├── vaultwarden  (password manager)
+│                                     ├── colony-psql  (shared PostgreSQL)
+│                                     ├── chatterbox   (Matrix Synapse + bridges)
+│                                     ├── jackflix     (media stack)
+│                                     ├── object       (MinIO, Harmonia Nix cache, Sharry, HedgeDoc, wastebin)
+│                                     ├── toot         (Bluesky PDS; Mastodon disabled)
+│                                     ├── waffletail   (Tailscale subnet router / exit node)
+│                                     ├── qclk         (WireGuard management appliance)
+│                                     ├── gam          (Terraria server)
+│                                     └── jam          (raw nspawn customer container)
+├── whale2 ─── podman/OCI host for game servers
+├── git ────── Gitea + Gitea Actions runner
+├── mail ───── Debian VM running mailcow (not NixOS)
+└── darts ──── third-party/customer VM (opaque, not NixOS)
+```
+
+## Site: home
+
+Redundant routers, VM host, storage, IoT containers and the workstation — see
+[`sites/home/README.md`](sites/home/README.md). The hand-configured switch fabric (jim/dave/brian)
+and the Digiweb WAN path are documented in [`sites/home/switches.md`](sites/home/switches.md).
+
+```
+h.nul.ie
+├── palace (physical VM host — AMD, 100G, SR-IOV)
+│   ├── river ── primary router VM (PPPoE / Digiweb WAN)
+│   ├── cellar ─ NVMe-oF / SPDK storage target VM
+│   └── sfh ──── container host VM ("services for home")
+│       ├── hass ── Home Assistant + Frigate + MQTT (container)
+│       └── unifi ─ UniFi controller (container)
+├── stream (physical secondary router — Virgin Media WAN)
+└── castle (workstation / gaming desktop — netboot, NVMe-oF root)
+```
+
+## Remote boxes
+
+Edge VPSes and the remote `kelder` site — see [`remote/README.md`](remote/README.md).
+
+- [`remote/britway.md`](remote/britway.md) — London VPS: headscale control plane, Tailscale exit
+  node, BGP edge.
+- [`remote/britnet.md`](remote/britnet.md) — Birmingham VPS: Tailscale exit node, WireGuard hub.
+- [`remote/kelder.md`](remote/kelder.md) — secondary home server: container host, Nextcloud,
+  Samba, WireGuard to `estuary`.
+  - [`remote/kelder-acquisition.md`](remote/kelder-acquisition.md) — media acquisition container.
+  - [`remote/kelder-spoder.md`](remote/kelder-spoder.md) — Nextcloud + nginx container.
+
+## Mobile boxes
+
+- [`mobile/tower.md`](mobile/tower.md) — laptop workstation.
+
+## Misc
+
+- [`misc/installer.md`](misc/installer.md) — the custom NixOS installer image.
+
+## A note on the assignment tables
+
+The "Network assignments" tables in box pages (between `<!-- assignments: … -->` markers) are
+**generated from the flake** (`nixos.allAssignments`) by `nix run .#update-docs-assignments` —
+CI refreshes them on push. Hand-written prose and the table's Notes column are preserved; don't
+hand-edit the other cells.
@@ -0,0 +1,242 @@
+# Architecture
+
+This flake does **not** use the stock pattern of calling `nixosSystem` once per host in
+[`flake.nix`](../flake.nix). Instead it runs a single `lib.evalModules` evaluation over its own
+module tree, producing one big top-level config (`self.nixfiles`) from which the real flake
+outputs (`nixosConfigurations`, `homeConfigurations`, `nixosModules`, `deploy`, …) are derived.
+Per-host files declare *options* (`nixos.systems.<name>`); the machinery in
+[`nixos/default.nix`](../nixos/default.nix) and
+[`home-manager/default.nix`](../home-manager/default.nix) turns those into evaluated NixOS /
+home-manager configurations.
+
+## The top-level evaluation
+
+`flake.nix`'s `outputs` builds a `nixfiles` attrset via `evalModules` over:
+
+- An inline module that seeds `_module.args` (`lib`, `pkgsFlakes`, `hmFlakes`, `self`, `inputs`,
+  `pkgs'`), sets `nixos.secretsPath = ./secrets`, and sets the global deploy-rs SSH option
+  `deploy-rs.deploy.sshOpts = [ "-i" ".keys/deploy.key" ]`.
+- `nixos/modules/misc/assertions.nix` from the unstable nixpkgs (so the top-level evaluation has
+  the standard `assertions` / `warnings` options).
+- [`nixos/`](../nixos/default.nix) — defines `nixos.*` options (`systems`, `modules`,
+  `allAssignments`, `vpns`, `secretsPath`) and `mkSystem`.
+- [`home-manager/`](../home-manager/default.nix) — defines `home-manager.*` options (`homes`,
+  `modules`) and `mkHome`.
+- [`deploy-rs.nix`](../deploy-rs.nix) — defines `deploy-rs.*` and renders the deploy config.
+- Every file in the `configs` list — the boxes themselves
+  ([`nixos/boxes/`](../nixos/boxes) plus [`nixos/installer.nix`](../nixos/installer.nix); the
+  home-manager entries are currently commented out, see [Home-manager](#home-manager)).
+
+The resulting `nixfiles.config` is mapped onto flake outputs:
+
+| Top-level option (`nixfiles.config`) | Flake output |
+|---|---|
+| `nixos.systems.<name>.rendered` | `nixosConfigurations.<name>` |
+| `home-manager.homes.<name>.configuration` | `homeConfigurations.<name>` |
+| `nixos.modules` | `nixosModules` |
+| `home-manager.modules` | `homeModules` |
+| `deploy-rs.rendered` | `deploy` |
+
+`nixfiles` itself is also a flake output, so anything in the top-level config can be addressed
+directly — e.g. `build-iso` builds
+`.#nixfiles.config.nixos.systems."<host>".configuration.config.my.buildAs.iso`. The flake also
+exposes `lib` (the extended unstable nixpkgs lib, including `lib.my`), `inputs`, `nixpkgs`
+(the `pkgs'` channel sets), and `overlays.default` (the custom packages from
+[`pkgs/`](../pkgs/default.nix)).
+
+Per platform (`eachDefaultSystem`), the flake produces:
+
+- `packages` — everything from `pkgs/`, flattened.
+- `checks` — every `homeConfigurations.*.activationPackage` plus deploy-rs's own `deployChecks`.
+- `devShells.default` — the `numtide/devshell` from [`devshell/`](../devshell) (see
+  [deployment.md](deployment.md)).
+- `ci.<system>` — one attr per buildable thing (`system-<name>`, `home-<name>` with `@` mangled
+  to `-at-`, `package-<name>`, plus `shell`), consumed by CI; `ciDrv` is a `linkFarm` of all of
+  them.
+
+## Systems: `systemOpts` and `mkSystem`
+
+Each entry of `nixos.systems` is a submodule defined by `systemOpts` in
+[`nixos/default.nix`](../nixos/default.nix):
+
+| Option | Type / default | Meaning |
+|---|---|---|
+| `system` | enum of `defaultSystems` | Nix platform string, e.g. `"x86_64-linux"`. |
+| `nixpkgs` | one of `unstable`/`stable`/`mine`/`mine-stable`, default `"unstable"` | nixpkgs channel for the system. |
+| `home-manager` | same enum, defaults to `nixpkgs` | home-manager channel. |
+| `hmNixpkgs` | same enum, defaults to `nixpkgs` | nixpkgs channel used for home-manager's own pkgs when it doesn't share the system's. |
+| `docCustom` | bool, default `false` | Include nixfiles' custom modules in the generated NixOS manual (slow). |
+| `assignments` | attrsOf `assignmentOpts` | The box's network assignments (see [networking.md](networking.md)). |
+| `extraAssignments` | attrsOf attrsOf `assignmentOpts` | Extra assignments for things not on the box itself (e.g. the routers' floating VIP entries). |
+| `configuration` | custom merged type | The actual NixOS configuration module(s); merging runs `mkSystem`. |
+| `rendered` | unspecified, default `configuration` | What ends up in `nixosConfigurations.<name>` — overridden for boxes built as something other than a plain system (e.g. `installer` renders `config.my.asISO`, containers render `config.my.asContainer`). |
+
+`mkSystem` does the real work:
+
+1. It selects the channel's nixpkgs flake (`pkgsFlakes.${config'.nixpkgs}`) and imports
+   `nixos/lib/eval-config.nix` **by hand**, because the flake force-sets `lib` and
+   `eval-config.nix` would otherwise import its own unextended one. The lib it passes is the
+   channel's `pkgs.lib` extended with `lib.my` plus a `versionOverlay` that stamps
+   `system.nixos` with the flake revision and channel.
+2. `specialArgs` gives every module access to `self`, `inputs`, `pkgsFlakes`, the box's own
+   `pkgsFlake`, `allAssignments` (every box's assignments), and `systems` (all of
+   `nixos.systems`). Passing these via `specialArgs` (rather than module `imports`) avoids
+   infinite recursion.
+3. The module list is: the home-manager channel's `nixosModules.default`, every module in
+   `nixos.modules` (the shared modules, see below), an inline module, and the box's own
+   `configuration` definitions (wrapped with `inlineModule'` so error messages keep file
+   provenance).
+4. The inline module wires the plumbing:
+   - `_module.args`: `secretsPath`, `vpns` (the flake-level `nixos.vpns`), the box's own
+     `assignments`, and `pkgs'` — an attrset of **all four nixpkgs channels** for this box's
+     platform, so a module can grab a package from another channel (e.g. `pkgs'.stable.foo`).
+   - `system.name` is the attribute name of the box.
+   - `networking.hostName` / `networking.domain` default from the `internal` assignment (see
+     [networking.md](networking.md)).
+   - `nixpkgs.system` is set and the overlays computed at flake level (lib overlay + custom
+     packages) are passed through, so `pkgs` is imported modularly with config and overlays
+     applied.
+   - home-manager integration: `useGlobalPkgs` defaults to true when the system and
+     home-manager channels match (a warning is emitted when they deliberately differ);
+     `sharedModules` is every module in `home-manager.modules` plus an inline module that
+     passes `pkgsPath` / `pkgs'` into home-manager, disables the release check, and pins
+     `home.stateVersion` via `homeStateVersion` (currently `22.11` for stable-flavoured
+     channels, `23.05` otherwise).
+5. Finally `applyAssertions` throws with all failed assertion messages (and shows warnings)
+   when the merged configuration is forced.
+
+## The four nixpkgs channels
+
+`flake.nix` builds `pkgsFlakes` (nixpkgs) and `hmFlakes` (home-manager):
+
+| Channel | nixpkgs input | home-manager input |
+|---|---|---|
+| `unstable` | `nixpkgs/nixos-unstable` | `home-manager` (master) |
+| `stable` | `nixpkgs/nixos-26.05` | `home-manager/release-26.05` |
+| `mine` | `github:devplayer0/nixpkgs/devplayer0` (personal fork) | alias of `unstable` (no fork exists) |
+| `mine-stable` | `github:devplayer0/nixpkgs/devplayer0-stable` | alias of `stable` |
+
+Each channel's `lib` is extended with the `libOverlay` (`lib.my` + flake-utils). Two package
+sets are built per channel:
+
+- `pkgs'` — with the devshell/ragenix/deploy-rs/home-manager overlays; used for the dev shell
+  and `packages`, and exposed as the flake's `nixpkgs` output.
+- `configPkgs'` — with just the lib + custom-packages overlays and an `allowUnfreePredicate`
+  (Widevine / Chromium); this is the `pkgs'` threaded into the top-level evaluation, from
+  which each box's `pkgs` and per-channel `pkgs'` module args (and home-manager's base
+  `pkgs`) are taken.
+
+A box picks its channel with `nixpkgs = "mine"` (most boxes), `"mine-stable"` (e.g. `colony`),
+etc. Inside a module, `pkgs` is the selected channel and `pkgs'` is the attrset of all four
+(`pkgs'.<channel>.<attr>`).
+
+## `lib.my` and the `my.*` namespace
+
+[`lib/default.nix`](../lib/default.nix) extends nixpkgs `lib` with a `my` attrset. The
+flake-level `lib` (`pkgsFlakes.unstable.lib` so extended) is for platform-independent flake
+use only; each box gets its own channel's lib extended the same way. Contents:
+
+- **Option helpers** — `mkOpt'`, `mkBoolOpt'`, `nullOrOpt'`, `mkDefault'` (`mkOverride 900`,
+  slightly stronger than `mkDefault`), `mkVMOverride'`, `inlineModule'` / `inlineModule`
+  (attach `_file` provenance), `commonOpts` (the shared `system`/`nixpkgs`/`home-manager`
+  options and the `moduleType` used for exported modules), `applyAssertions`, `duplicates`.
+- **`lib.my.net`** — CIDR/IP math (`net.cidr.host`, `net.cidr.subnet`, `net.types.ipv4`, …)
+  from the `libnetRepo` input (`oddlama/nixos-extra-modules`' `netu.nix`). Used for nearly all
+  address arithmetic; addresses are almost never written literally.
+- **`lib.my.c`** — shared constants from [`lib/constants.nix`](../lib/constants.nix): static
+  UID/GID assignments, kernel package selection (`kernel.lts`/`kernel.latest`), nginx config
+  snippets, networkd snippets (`networkd.noL3`), the binary-cache settings (`nix.cache`),
+  and the per-site domains/prefixes/VIPs described in [networking.md](networking.md).
+- **`lib.my.dns`** — zone-generation helpers from [`lib/dns.nix`](../lib/dns.nix): `fwdRecords`
+  / `ptrRecords` / `ptr6Records` rendered from `allAssignments`, plus the LUA-record helpers
+  (`ifaceA`, `lookupIP`) used by the home routers.
+- **networkd helpers** — `networkdAssignment` (assignment → `systemd.network` network, see
+  [networking.md](networking.md)), `mkVLAN`, `dockerNetAssignment`.
+- **Misc** — `mkDefaultSystemsPkgs`, `flakePackageOverlay` (wrap another flake's package as an
+  overlay), `isIPv6` / `parseIPPort` / `netBroadcast`, `nft` chain-name helpers,
+  `systemdAwaitPostgres`, `vm.*` LVM disk descriptors for the `vms` module, the typed
+  `deploy-rs` option definitions (`lib.my.deploy-rs`), `netbootKeaClientClasses` (iPXE/EFI
+  DHCP client classes), and `homeStateVersion`.
+
+Custom NixOS / home-manager modules declare their options under the `my.*` namespace (the root
+`options.my` is declared in `nixos/modules/common.nix` / `home-manager/modules/common.nix`) —
+e.g. `my.secrets`, `my.build`, `my.tmproot`, `my.firewall`, `my.server`, `my.deploy`,
+`my.vms`, `my.containers`.
+
+## Shared modules
+
+Registered in [`nixos/modules/_list.nix`](../nixos/modules/_list.nix) and applied to **every**
+system (box files opt in per-feature via `my.*` options):
+
+| Module | Provides |
+|---|---|
+| `common` | Baseline for all boxes: imports the impermanence, ragenix (age), sharry, copyparty and harmonia NixOS modules; pins `system.stateVersion`; `doas` instead of `sudo`; immutable users; nix settings (flakes, `ca-derivations`, the `nix-cache.nul.ie` substituter); declares the `my` option root. |
+| `user` | `my.user` — the primary user: `users.users` + matching `home-manager.users` entry, wheel/doas, SSH authorized key from `.keys/me.pub`, shell taken from the home config, home persistence under tmproot. |
+| `build` | `my.build` — alternate build targets via `extendModules`: `my.buildAs.devVM` (QEMU dev VM), `iso`, `container`, `kexecTree`, `netbootTree`/`netbootArchive`; `my.build.isDevVM` marker; `allHardware` profile toggle. |
+| `dynamic-motd` | `my.dynamic-motd` — runs a script via `pam_exec` to generate the MOTD on login/ssh. |
+| `tmproot` | `my.tmproot` — tmpfs `/` plus impermanence persistence (`persistence.dir`), with a `tmproot-unsaved` helper that walks the root tmpfs and lists files not covered by persistence. |
+| `firewall` | `my.firewall` — nftables firewall: `tcp.allowed`/`udp.allowed` port lists, `trustedInterfaces`, NAT with `forwardPorts`, `extraRules` escape hatch; sane ICMP/ICMPv6/PMTUD rules baked in. |
+| `server` | `my.server.enable` — server commonalities: getty autologin, LLMNR off, tightened fstrim timers, GUI and NixOS documentation off. |
+| `deploy-rs` | `my.deploy` — per-box deploy-rs node/profile generation and the `deploy` user; see [deployment.md](deployment.md). |
+| `secrets` | `my.secrets` — ragenix/agenix wiring: `files` to decrypt from `secrets/`, host `key` to encrypt for, identity paths derived from the OpenSSH host keys (persistence-aware), dev-key identity inside dev VMs. |
+| `containers` | `my.containers.instances` — `systemd-nspawn` NixOS containers whose systems come from `nixos.systems` rendered `asContainer`, with bridge/macvlan networking, bind mounts, `hotReload` (reload instead of reboot) and a placeholder "dummy" init before first deploy. |
+| `vms` | `my.vms.instances` — QEMU/KVM VMs as systemd units: TAP/bridge networking, LVM-backed disks, VFIO host-device passthrough (with udev tagging), UEFI, SPICE/TTY/QMP unix sockets under `/run/vms/`, clean shutdown via QMP `system_powerdown`. |
+| `network` | Baseline networking: networkd only (`useDHCP = false`), IPv6 on, resolved domain/negative-cache settings; dev VMs get DHCP on `eth0` and an SSH port forward. |
+| `pdns` | `my.pdns` — PowerDNS: authoritative server driven by BIND-style zone files (with templating and serial management) and recursor extra-settings wiring. |
+| `nginx-sso` | `my.nginx-sso` — runs `nginx-sso` (instead of the stock NixOS module) and generates per-instance nginx `auth_request` include files. |
+| `gui` | `my.gui.enable` (default on) — desktop baseline: graphics, polkit, swaylock PAM entry, Android udev rules, screenshot tmpdir. |
+| `l2mesh` | Consumes `nixos.vpns.l2` — builds the VXLAN + IPsec layer-2 meshes between edge routers; see [networking.md](networking.md). |
+| `borgthin` | `my.borgthin.jobs` — borg backups of thin-LVM snapshots to local or SSH repos on a systemd timer, with pruning. |
+| `nvme` | `my.nvme` — sets the NVMe host NQN/hostid and, optionally, NVMe-oF-over-RDMA boot in the initrd. |
+| `spdk` | `my.spdk` — SPDK target configuration (JSON RPC-driven) plus `spdk-rpc`/`spdk-setup`/`spdk-debug` helper tools. |
+| `librespeed` | `my.librespeed` — a LibreSpeed speedtest: generated frontend (from the `librespeed-go` package in `pkgs/`) plus backend settings. |
+| `netboot` | `my.netboot` — iPXE netboot server (TFTP/HTTP, menu) and the client-side loader that installs boot entries. |
+
+Home-manager modules, registered in
+[`home-manager/modules/_list.nix`](../home-manager/modules/_list.nix) and applied to every
+home (including the per-user homes attached to systems via `my.user.homeConfig`):
+
+| Module | Provides |
+|---|---|
+| `common` | Home baseline: `my.shell`, `my.ssh.authKeys`, `my.isStandalone` (home-manager running standalone vs inside NixOS), fish setup and completions generation, common programs. |
+| `gui` | `my.gui` — the sway-based desktop: waybar, notifications, lock/saver plumbing (including the "doomsaver"), terminal and fonts. |
+| `deploy-rs` | `my.deploy` for standalone homes — generates a `home` profile (home-manager activation) and deploy node; auto-disabled for NixOS-attached homes. |
+| `swaync` | `my.swaync` — typed configuration module for Sway Notification Center. |
+
+## Adding a box
+
+1. Create a file (or directory with a `default.nix`) under [`nixos/boxes/`](../nixos/boxes)
+   that sets `nixos.systems.<name> = { system = "x86_64-linux"; nixpkgs = "mine"; assignments =
+   { … }; configuration = { … }: { … }; };`. Nested boxes (VMs, containers) live under their
+   host's directory (e.g. `nixos/boxes/colony/vms/estuary/`).
+2. Add the path to the `configs` list in [`flake.nix`](../flake.nix).
+3. Give the box an `internal` assignment (name/domain) if it gets one — `hostName`/`domain`
+   default from it — and declare `my.secrets.key` if it has secrets.
+4. Evaluate it with `check-system <name>` (cheap) before `build-system <name>`.
+
+## Adding a shared module
+
+1. Drop the file in [`nixos/modules/`](../nixos/modules) (or
+   [`home-manager/modules/`](../home-manager/modules)) with options under `my.*`, declared via
+   the `lib.my` helpers (`mkOpt'`, `mkBoolOpt'`).
+2. Register it in the corresponding `_list.nix` (name → path). It is then applied to every
+   box and exported as `nixosModules.<name>` / `homeModules.<name>`.
+
+## Home-manager
+
+Home-manager has the same two-level structure: `homeOpts` (in
+[`home-manager/default.nix`](../home-manager/default.nix)) defines `system`, `nixpkgs`,
+`home-manager`, `homeDirectory`, `username`, and `configuration`; `mkHome` calls the selected
+channel's `lib.homeManagerConfiguration`, passing the channel's `pkgs'` (with `config` emptied
+so home-manager applies overlays/config modularly itself), `extraSpecialArgs` (`inputs`,
+`pkgsFlakes`, `pkgsFlake`), every module in `home-manager.modules`, an inline module that sets
+`home.homeDirectory` / `home.username` and provides `pkgs'` (all channels) as a module arg,
+and the pinned `home.stateVersion`.
+
+Standalone homes live in [`home-manager/configs/`](../home-manager/configs) and are named
+`<user>@<host>` (deploy-rs mangles the `@` to `-at-`). `macsimum` is a home-manager-only
+config — an `x86_64-darwin` macOS box with no NixOS side at all. Note that **no homes are
+currently wired into the flake**: the `home-manager/configs/macsimum.nix` entry in `configs` is
+commented out, and `home-manager/configs/castle.nix` exists on disk but is not listed either,
+so `nixfiles.config.home-manager.homes` currently evaluates to `{}`. Most day-to-day user
+configuration instead rides along with systems through `my.user.homeConfig`.
@@ -0,0 +1,178 @@
+# Deployment and workflows
+
+How boxes get deployed, the devshell commands that drive everyday work, secrets, dev VMs, and
+CI.
+
+## deploy-rs
+
+### Rendering
+
+The top-level [`deploy-rs.nix`](../deploy-rs.nix) renders the flake's `deploy` output:
+
+- It collects `nixos.systems` and `home-manager.homes` (mangling `@` in home names to
+  `-at-`, since deploy-rs node names can't contain `@`). An assertion rejects name collisions
+  between systems and homes.
+- A system/home becomes a deploy node only when `configuration.config.my.deploy.enable` is
+  true. The node config is the *definitions* of the box's `my.deploy.node` option, re-imported
+  as a module so deploy-rs's submodule merging still applies.
+- `autoRollback` and `magicRollback` are off globally; global `sshOpts` are
+  `[ "-i" ".keys/deploy.key" ]` (set in [`flake.nix`](../flake.nix)).
+- The result is passed through `lib.my.deploy-rs.filterOpts`, which strips nulls so unset
+  options don't override deploy-rs defaults, and exposed as `deploy = deploy-rs.rendered`.
+
+### Per-box nodes
+
+The shared [`nixos/modules/deploy-rs.nix`](../nixos/modules/deploy-rs.nix) module provides
+`my.deploy`:
+
+- `my.deploy.enable` defaults to true, but is **automatically forced off for dev VMs and
+  NixOS containers** (`my.build.isDevVM` / `boot.isContainer` — containers are deployed
+  through their host instead, see below). The installer also disables it explicitly.
+- Node defaults: `hostname` = the box's FQDN, `sshUser = "deploy"`, `user = "root"`,
+  `sudo = "doas -u"` (or `sudo -u`), `sshOpts` = the box's first OpenSSH port. The module
+  creates the `deploy` system user (bash shell, wheel, keys from `my.deploy.authorizedKeys`,
+  which defaults to `.keys/deploy.pub`).
+- Generated profiles:
+  - **`system`** — activates `config.system.build.toplevel` with
+    `switch-to-configuration switch`, with workarounds for the `/tmp` cwd issue
+    (NixOS/nixpkgs#73404) and the systemd-boot `loader.conf` default-entry issue
+    (deploy-rs#31), then prunes old generations (`nix-env --delete-generations +10`,
+    tunable via `keepGenerations`, 0 disables).
+  - **`container-<name>`** — one per `my.containers.instances` entry, activating the
+    container's `my.buildAs.container` toplevel into
+    `/nix/var/nix/profiles/per-container/<name>/system`. With `hotReload` (default) the
+    profile *reloads* `systemd-nspawn@<name>` (restarting only if the container is down or
+    still running the placeholder "dummy" init); otherwise it restarts it. Generation cleanup
+    applies here too.
+  - Profiles deploy in order `system`, then the containers.
+
+### Usage
+
+- `deploy .#<host>` — the devshell's `deploy` is a wrapper that adds `--skip-checks`. Node
+  names are the system names (`deploy .#git`).
+- `deploy-multi <hosts...>` — loops `deploy` over several nodes (extra args via `$O`).
+- `deploy --boot .#<host>` — stages the config as the boot default **without** live-switching
+  (deploy-rs's `--boot` maps to the custom activation's `boot` phase, which also does the
+  generation cleanup). Use it when a live `switch` would cut the box off mid-change (e.g. a
+  router WAN rework), then reboot to cut over.
+- `ssh-machine <name> [cmd]` — resolves `user@host` and the merged global+node `sshOpts` from
+  `.#deploy` (same `my.deploy.enable` gate as `deploy`, same `@` → `-at-` mangling), then
+  execs `ssh`. Boxes default to the `fish` login shell; pipe multi-statement remote scripts
+  through `bash` (`ssh-machine <name> bash -s < script.sh`). If outbound SSH hangs at the
+  publickey step (flaky agent), use `SSH_AUTH_SOCK= ssh-machine …`.
+- `nix flake check` includes deploy-rs's own `deployChecks` for the whole `deploy` attrset.
+
+## Devshell commands
+
+The repo ships a `numtide/devshell` ([`devshell/`](../devshell), entered via `direnv`). Run a
+command with no arguments for its help. From
+[`devshell/commands.nix`](../devshell/commands.nix):
+
+| Command | What it does |
+|---|---|
+| `check-system <host> [nix args]` | Evaluates `.#nixosConfigurations."<host>".config.system.build.toplevel.drvPath` — catches module/option errors without building. Prefer this to validate a config change. |
+| `build-system <host> [nix args]` | Builds the system's `toplevel` (extra args pass through to `nix build`). |
+| `build-n-switch <args>` | `doas nixos-rebuild --flake .` (adds the repo as a git `safe.directory` for root first). |
+| `build-home <name> [nix args]` | Builds `.#homeConfigurations."<name>".activationPackage`. |
+| `home-switch [args]` | `home-manager switch --flake .`. |
+| `deploy [args]` | `deploy-rs --skip-checks` (wrapper package in `devshell/default.nix`). |
+| `deploy-multi <nodes...>` | Deploys several nodes in sequence. |
+| `ssh-machine <name> [cmd]` | SSH to a system or home by name, resolving target/options from its deploy-rs node (see above). |
+| `run-vm <host>` | Boots a system as a dev VM: installs `.keys/dev.key` into a temp `xchg/`, then `nix run`s `config.my.buildAs.devVM`. |
+| `build-iso <host>` | Builds `config.my.buildAs.iso`. |
+| `build-kexec <host>` | Builds `config.my.buildAs.kexecTree`. |
+| `build-netboot <host>` | Builds `config.my.buildAs.netbootTree`. |
+| `ragenix [args]` | `ragenix --identity .keys/dev.key` (see [Secrets](#secrets)). |
+| `repl` | `nix repl .#`. |
+| `fmt [args]` | `nixpkgs-fmt` (the canonical formatter). |
+| `update-nixpkgs` | `nix flake update nixpkgs-{unstable,stable,mine,mine-stable}`. |
+| `update-home-manager` | `nix flake update home-manager-{unstable,stable}`. |
+| `update-installer` | Force-pushes the `installer` tag to trigger the installer release workflow. |
+| `home-link` / `home-unlink` | Symlink (or remove) this `flake.nix` at `~/.config/home-manager/flake.nix` for standalone `home-manager` use. |
+| `qemu-genmac` | Prints a random QEMU-suitable MAC (`52:54:00:xx:xx:xx`). |
+| `ssh-get-ed25519 <host>` | Prints a host's ed25519 pubkey via `ssh-keyscan`. |
+| `json2nix` | Converts JSON on stdin to formatted Nix. |
+
+From [`devshell/install.nix`](../devshell/install.nix) (driven by `$INSTALLER`, the address of
+a running custom installer, SSHing as root with `.keys/deploy.key`):
+
+| Command | What it does |
+|---|---|
+| `installer-shell [cmd]` | Runs a command (default: a shell) inside the installer. |
+| `do-install [--no-bootloader] [--no-substitute] <system>` | Builds the system's `toplevel`, `nix copy`s it into the installer's target store, sets the system profile, and activates it with `switch-to-configuration boot` (with `NIXOS_INSTALL_BOOTLOADER=1` unless `--no-bootloader`). |
+
+From [`devshell/vm-tasks.nix`](../devshell/vm-tasks.nix) (remote VM consoles; they forward the
+VM's unix sockets from `/run/vms/<vm>/` on `<host>` over SSH):
+
+| Command | What it does |
+|---|---|
+| `vm-tty <host> <vm>` | Serial TTY of a VM in `minicom`. |
+| `vm-monitor <host> <vm>` | QEMU monitor socket in `minicom`. |
+| `vm-viewer <host> <vm>` | SPICE display in `virt-viewer` (not on Darwin). |
+
+## Secrets
+
+Secrets are age-encrypted files in [`secrets/`](../secrets), managed with **ragenix** (a fork
+with a rekey flag, from the flake inputs).
+
+- Each box declares `my.secrets.key` (the host public key its secrets encrypt to) and
+  `my.secrets.files.<name>` (files to decrypt; attribute values merge into the agenix secret,
+  e.g. `owner`/`mode`). At runtime the `secrets` module decrypts via
+  `age.secrets."<name>".path`, with identity paths derived from the box's OpenSSH host keys —
+  read from the tmproot persistence dir when there is one, since agenix runs before persisted
+  keys would otherwise be available.
+- [`secrets.nix`](../secrets.nix) (the ragenix rules file at the repo root) is **generated**:
+  it evaluates the flake, collects every system's `my.secrets.files` + `key`, and emits each
+  secret with its recipient list — always including `.keys/dev.pub` so the dev key can open
+  everything. Re-running `ragenix -r` re-keys after adding a box or secret.
+- The `ragenix` devshell command wraps `ragenix --identity .keys/dev.key`; use it to
+  edit/rekey secrets.
+- The `.keys/` directory holds the dev key (`dev.key`/`dev.pub`), the deploy key
+  (`deploy.key`/`deploy.pub`, authorized on every box's `deploy` user), and assorted other
+  keys referenced by `lib.my.c.sshKeyFiles`. It is required for editing secrets, deploying,
+  and running dev VMs.
+
+## Dev VMs
+
+Any system can be built as a throwaway QEMU VM via `config.my.buildAs.devVM` (the `build`
+module extends the config with `qemu-vm.nix` and sets `my.build.isDevVM`). `run-vm <host>`
+creates a temp dir, installs `.keys/dev.key` as `xchg/dev.key`, and runs the VM; inside, the
+`secrets` module switches `age.identityPaths` to that dev key (`my.secrets.vmKeyPath`,
+default `/tmp/xchg/dev.key`), so dev VMs can decrypt the boxes' secrets without the real host
+keys. Dev VMs also get DHCP on `eth0`, an SSH port forward (host 2222 → guest 22), and are
+automatically excluded from deploy targets.
+
+## CI
+
+GitHub/Gitea Actions workflows live in [`.gitea/workflows/`](../.gitea/workflows):
+
+- **`ci.yaml`** (push to `master`): `nix flake check --no-build`, then for every attribute of
+  `.#ci.x86_64-linux` (each system as `system-<name>`, each home as `home-<name>` with `@` →
+  `-at-`, each package as `package-<name>`, plus the dev `shell`) it builds and pushes the
+  result to the Harmonia binary cache with [`ci/push-to-cache.sh`](../ci/push-to-cache.sh).
+  Finally it builds `.#ciDrv.x86_64-linux` (a `linkFarm` of all CI attrs) and pushes it with
+  `UPDATE_PROFILE=1`, which updates the `nixfiles` profile on the cache box and garbage
+  collects paths older than 60 days. The cache is `ssh-ng://harmonia@object-ctr.ams1.int.nul.ie`
+  (remote store `/var/lib/harmonia`), keyed by the `HARMONIA_SSH_KEY` secret with a pinned
+  `ci/known_hosts`; clients consume it as `https://nix-cache.nul.ie` (see `lib.my.c.nix.cache`).
+- **`installer.yaml`** (push of the `installer` tag; `update-installer` refreshes it): builds
+  the installer's `my.buildAs.iso` and `my.buildAs.netbootArchive` and attaches both to a
+  release.
+- **`update-docs.yaml`** (push to the docs branch, skipping its own commits): runs
+  `nix run .#update-docs-assignments` and commits any changes as
+  `docs: update assignment tables`.
+
+### The docs assignment-table updater
+
+`update-docs-assignments` is registered in [`pkgs/default.nix`](../pkgs/default.nix) (a
+`writeShellScriptBin` wrapping [`ci/update-docs-assignments.py`](../ci/update-docs-assignments.py)).
+It evaluates `.#nixfiles.config.nixos.allAssignments` to JSON, walks `docs/**/*.md`, and for
+each file looks for a marker line `<!-- assignments: <box> -->` where `<box>` is the **file's
+own stem** (e.g. `river.md` → `river`) and names a real box. The generated table (columns
+`Name | Assignment | IPv4 | IPv6 | Domain | Notes`) is (re)written between the marker and a
+closing `<!-- assignments-end -->` line, with an `<!-- assignments-start -->` line inserted
+after the marker on the first run. Hand-written text in the **Notes** column is preserved
+across runs (keyed by the Assignment cell), so notes survive regeneration. Pages that want a
+table only need the marker + end marker; pages without markers (including the three top-level
+docs) are left alone. The script exits non-zero when it changed something, which is how the
+workflow knows whether to commit.
@@ -0,0 +1,42 @@
+# installer
+
+The custom NixOS installer image used to bootstrap new boxes.
+
+- **Source:** [`nixos/installer.nix`](../../nixos/installer.nix)
+- **Host:** — (a build target, not a deployed box)
+
+## Role
+
+- Defines `nixos.systems.installer`, a minimal server system rendered as a bootable ISO via
+  `config.my.buildAs.iso` (`my.asISO`); the same base can also be built as a kexec or netboot
+  tree.
+- Build it with the devshell commands: `build-iso installer` (or `build-kexec installer` /
+  `build-netboot installer`). The `update-installer` command force-tags `installer` to
+  trigger a release rebuild in CI.
+
+## Image contents
+
+- Broad hardware support (`my.build.allHardware` pulls in the nixpkgs all-hardware profile);
+  EFI- and USB-bootable, zstd-compressed squashfs. Volume ID
+  `jackos-<release>-<arch>`, menu label "/dev/player0 Installer", image base name
+  `jackos-installer`.
+- Root SSH with the deploy key authorized (`PermitRootLogin prohibit-password`); a random
+  `installer-<hex>` hostname is set at boot.
+- `INSTALL_ROOT=/mnt` in the session environment, plus a `show-hw-config` alias wrapping
+  `nixos-generate-config --show-hardware-config --root $INSTALL_ROOT`.
+- NixOS documentation enabled, `wpa_supplicant` available but not started, aggressive GC
+  (`GC_INITIAL_HEAP_SIZE=1M`) and `vm.overcommit_memory=1` for low-memory targets, LVM thin
+  and NFS support.
+- No regular user (`my.user.enable = false`), no tmpfs-root management, no NAT, and not a
+  deploy target (`my.deploy.enable = false`).
+
+## Installing a box
+
+The devshell's installer commands ([`devshell/install.nix`](../../devshell/install.nix)) drive
+an install over SSH against a booted installer reachable at `$INSTALLER`:
+
+- `installer-shell` — get a shell on the installer.
+- `do-install <system>` — builds the system's toplevel, `nix copy`s the closure to the
+  installer's `$INSTALL_ROOT` remote store, sets the system profile, touches `/etc/NIXOS`,
+  and runs `switch-to-configuration boot` with `NIXOS_INSTALL_BOOTLOADER=1` (skip the
+  bootloader with `--no-bootloader`, skip substitution with `--no-substitute`).
@@ -0,0 +1,9 @@
+# Mobile boxes
+
+Portable workstations that move between networks — currently just `tower`. Mobile boxes have
+no static network assignments; they use DHCP/NetworkManager and reach other boxes over the
+tailnet.
+
+| Box | What it is | Docs |
+| --- | --- | --- |
+| `tower` | Framework Laptop 13 workstation | [tower.md](tower.md) |
@@ -0,0 +1,38 @@
+# tower
+
+Portable workstation — a Framework Laptop 13 (Intel), running the full GUI environment.
+
+- **Source:** [`nixos/boxes/tower/default.nix`](../../nixos/boxes/tower/default.nix)
+- **Host:** physical (laptop)
+
+## Role
+
+- Personal portable workstation: `my.gui.enable`, with Sway managed by home-manager.
+- Joins the tailnet through the headscale on [`britway`](../remote/britway.md) (fish abbr
+  `tsup` = `doas tailscale up --login-server=https://hs.nul.ie --accept-routes`).
+
+## Hardware / platform
+
+- Intel platform: microcode updates, `kvm-intel`, `intel_iommu=on`, `intel-media-driver` for
+  graphics; latest kernel (`lib.my.c.kernel.latest`).
+- Thunderbolt security (`bolt`), Bluetooth (`blueman` + tray applet), fingerprint reader
+  (`fprintd`; `doas` persists auth for `wheel`).
+- `tlp` power management, including battery charge thresholds (start 90% / stop 97%).
+
+## Storage
+
+- Two LUKS-encrypted partitions, `persist` and `home` (both `allowDiscards`); `/nix` is a
+  separate ext4 filesystem, `/boot` is vfat.
+- Persistent `/home` (`my.user.tmphome = false`) with an 8G tmpfs root
+  (`my.tmproot.size = "8G"`).
+
+## Networking / services
+
+- NetworkManager (`wpa_supplicant` backend) with `systemd-resolved`; networkd `wait-online`
+  is disabled. The Wi-Fi interface is renamed to `wifi` by MAC.
+- Steam and Wireshark enabled; `fstrim`, LVM thin provisioning.
+- `nix.gc.automatic = false` — GC is run manually on the laptop.
+
+## Notable config files
+
+- [`nixos/boxes/tower/default.nix`](../../nixos/boxes/tower/default.nix) — the whole box (single file).
@@ -0,0 +1,221 @@
+# Networking
+
+This page describes how addressing works across the boxes: the assignment mechanism, the
+per-site domains and prefixes, the home router HA pair, and the overlays/tunnels that tie the
+sites together. Switch-level home topology (jim/dave/brian, the ONT path) lives in
+[sites/home/switches.md](sites/home/switches.md).
+
+## Assignments
+
+Every box declares `nixos.systems.<name>.assignments`, an attrset of *assignments* (one per
+network the box is attached to). The option shape (`assignmentOpts` in
+[`nixos/default.nix`](../nixos/default.nix)):
+
+- `name` (defaults to the attribute name) and `altNames` — DNS names for the assignment.
+- `visible` (default `true`) — whether DNS helpers include it.
+- `domain` — DNS suffix for this assignment.
+- `mtu` — interface MTU (applied via the network's `linkConfig.MTUBytes`).
+- `ipv4.address` / `ipv4.mask` (default 24) / `ipv4.gateway` (defaults to host 1 of the
+  prefix; set explicitly to `null` when there is no gateway) / `ipv4.genPTR`.
+- `ipv6.address` (nullable — an assignment can be v4-only) / `ipv6.mask` (default 64) /
+  `ipv6.iid` (SLAAC static token instead of a full address) / `ipv6.gateway` / `ipv6.genPTR`.
+
+`extraAssignments` is a second, nested level for addresses that belong *to* a network but not
+to any single box — the home routers use it for their floating VIP entries (`router-hi`,
+`router-lo`, `router-ut`).
+
+All assignments are aggregated into `nixos.allAssignments` — every system's `assignments`
+merged with every system's `extraAssignments` — and passed to every module as the
+`allAssignments` argument, so any box can route to any other box's addresses without
+hardcoding. A flake-wide assertion fails evaluation if any IPv4 or IPv6 address appears in
+more than one assignment. Each box also receives its own assignments as the `assignments`
+module argument.
+
+Two pieces of machinery consume assignments:
+
+- `lib.my.networkdAssignment` ([`lib/default.nix`](../lib/default.nix)) renders an assignment
+  as a `systemd.network` network: static `address`/`gateway`, MTU, LLDP, and IPv6 RA handling
+  (`IPv6AcceptRA` when there's no static gateway or a static `iid` is set, with
+  `Token = static:<iid>`).
+- `mkSystem` defaults `networking.hostName` to `assignments.internal.name` (falling back to
+  the system name) and `networking.domain` to `assignments.internal.domain`. The shared
+  `network` module sets a fallback domain of `int.nul.ie` for boxes without one.
+
+## Domains
+
+The public domain is `nul.ie` (`lib.my.c.pubDomain`). Each site has its own internal domain
+(constants in [`lib/constants.nix`](../lib/constants.nix)):
+
+| Site | Domain |
+|---|---|
+| colony | `ams1.int.nul.ie` |
+| home | `h.nul.ie` |
+| britway | `lon1.int.nul.ie` |
+| britnet | `bhx1.int.nul.ie` |
+| kelder | `hentai.engineer` |
+
+## colony
+
+The colony box is a hosted server in Amsterdam (`ams1`); its public edge is the `estuary` VM
+(`94.142.240.44`, `2a02:898:0:20::329:1`), which NATs and filters for everything behind it.
+The internal prefixes (`lib.my.c.colony.prefixes`) are carved from `10.100.0.0/16` and
+`2a0e:97c0:4d2:10::/60`:
+
+| Network | IPv4 | IPv6 | Purpose |
+|---|---|---|---|
+| `base` | `10.100.0.0/24` | `2a0e:97c0:4d2:10::/64` | Base/management LAN (bridge on the host; estuary is `.1`) |
+| `vms` | `10.100.1.0/24` | `2a0e:97c0:4d2:11::/64` | VM network (host is `.1`, hands out RAs) |
+| `ctrs` | `10.100.2.0/24` | `2a0e:97c0:4d2:12::/64` | `systemd-nspawn` containers on the `shill` VM |
+| `oci` | `10.100.3.0/24` | `2a0e:97c0:4d2:13::/64` | Podman/OCI workloads on the `whale2` VM |
+| `qclk` | `10.100.4.0/24` | — | WireGuard endpoint instances in the `qclk` container |
+
+On top of that: `p2pTunnels` (`10.100.5.0/24`) holds point-to-point tunnel /30s (see
+[WireGuard tunnels](#wireguard-point-to-point-tunnels)); the `as211024` mesh gets
+`10.100.50.0/24` + `2a0e:97c0:4df::/64` (see [the L2 mesh](#the-as211024-l2-mesh)); and the
+`cust` block (`10.100.100.0/24`, `2a0e:97c0:4d2:2000::/56`) plus the `vip1`/`vip2`/`vip3`
+public blocks and the per-customer `mail` / `darts` / `jam` prefixes carry customer-facing
+services with their own public addresses (announced by BGP, routed via the host).
+
+## home
+
+The home site prefixes (`lib.my.c.home.prefixes`) come from `192.168.64.0/18` and
+`2a0e:97c0:4d0::/60`, with VLAN IDs from `lib.my.c.home.vlans`:
+
+| Network | VLAN | IPv4 | IPv6 | MTU | Purpose |
+|---|---|---|---|---|---|
+| `core` | — (macvlan) | `192.168.64.0/24` | — | 1500 | Router-to-router/core link |
+| `hi` | 100 | `192.168.68.0/22` | `2a0e:97c0:4d0:1::/64` | 9000 | High-speed LAN (jumbo frames) |
+| `lo` | 110 | `192.168.72.0/21` | `2a0e:97c0:4d0:2::/64` | 1500 | General LAN |
+| `untrusted` | 120 | `192.168.80.0/24` | `2a0e:97c0:4d0:3::/64` | 1500 | Untrusted / IoT |
+| `modem` | 130 (`wan`) | `192.168.0.0/24` | — | — | Virgin Media modem management (stream) |
+| `ont` | 140 (`wan-pon-ont`) | `192.168.100.0/24` | — | — | Digiweb ONT management (river) |
+
+Two more WAN-side VLANs exist: `pon-isp` (10), the ISP VLAN Digiweb delivers at the ONT and
+which is trunked untranslated to river, and `wan-pon-isp` (141), reserved for a future
+multi-ONT translation scheme — see
+[sites/home/switches.md](sites/home/switches.md) for the fabric side.
+
+The routers themselves (`river` = host 1, `stream` = host 2 in each prefix) are built from one
+definition, [`nixos/boxes/home/routing-common`](../nixos/boxes/home/routing-common/default.nix),
+parameterised by an index (`0` = river, `1` = stream) that derives per-box addresses, DHCP
+pool splits, VRRP state/priority and DNS `ns` numbering.
+
+### Router VIPs
+
+Clients never use a router's real address: each client VLAN has a floating VIP
+(`lib.my.c.home.vips`) that follows the VRRP master. The VIPs are also declared as
+`extraAssignments` (`router-hi`/`router-lo`/`router-ut`) so they appear in `allAssignments`
+and DNS:
+
+| Assignment | IPv4 | IPv6 |
+|---|---|---|
+| `router-hi` | `192.168.71.254/22` | `2a0e:97c0:4d0:1::ffff/64` |
+| `router-lo` | `192.168.79.254/21` | `2a0e:97c0:4d0:2::ffff/64` |
+| `router-ut` | `192.168.80.254/24` | `2a0e:97c0:4d0:3::ffff/64` |
+
+There is also a mesh-side VIP (`as211024`): `10.100.50.4` and `2a0e:97c0:4df:0:1::ffff`,
+which the other sites use as their next-hop into the home prefixes.
+
+### Router HA
+
+- **VRRP via `keepalived`** ([`routing-common/keepalived.nix`](../nixos/boxes/home/routing-common/keepalived.nix)):
+  separate v4/v6 instances (router IDs 51/52) on the `lan-core` link; index 0 (river) starts
+  as `MASTER`, priorities are `255 - index`. Track scripts ping public v4/v6 anycast targets
+  and demote a router whose WAN is down. All VIPs of a family move together.
+- **Clients get the VIP as gateway *and* DNS.** `kea` hands out `vips.<vlan>.v4` as both
+  `routers` and `domain-name-servers`, with the two routers serving disjoint pool halves;
+  `radvd` advertises the v6 VIP as RDNSS (the `untrusted` VLAN gets Cloudflare instead) and
+  is started/stopped by keepalived's `notify_master`/`notify_backup` hooks so only the master
+  sends RAs.
+- **`pdns-recursor` binds the VIPs directly** ([`routing-common/dns.nix`](../nixos/boxes/home/routing-common/dns.nix)),
+  with `net.ipv4.ip_nonlocal_bind` / `net.ipv6.ip_nonlocal_bind` so the backup can listen on
+  addresses it doesn't currently hold — failover doesn't depend on client resolver timeouts.
+  The recursor forwards the site's zones to the local authoritative PowerDNS on `127.0.0.1:5353`.
+- **`wan-online.target`** is a shared, initially-inert systemd target meaning "the public
+  IPv4 WAN route is up". `routing-common` only declares it; each box wires how it's reached —
+  stream gates it on a oneshot that waits for the DHCP default route on `wan`, river's `pppd`
+  `ip-up`/`ip-down` hooks start/stop it. Consumers (e.g. `ipsec`, the RA-default-route
+  cleanup) attach **to** it with `wantedBy` + `partOf` + `after`, never `requires`/`wants`,
+  so the target is never pulled in early and services re-load on WAN flap.
+
+### WAN paths (summary)
+
+- **river** (a VM on `palace`): PPPoE to Digiweb via `services.pppd`, running directly on the
+  ISP's VLAN 10 (`wan-pon-isp`, trunked untranslated through the switches; baby-jumbo MTU 1508
+  so the PPP session is a clean 1500). The static IP is requested in IPCP; the pppd hooks own
+  `wan-online.target`. The ONT's management subnet (`192.168.100.0/24`) is reached on
+  `wan-pon-ont` (VLAN 140, PVID'd at the brian switch), where river takes `.100`.
+- **stream** (bare metal): DHCP on the Virgin Media cable modem (VLAN 130) on `wan`, with a
+  static modem-management address (`192.168.0.100/24`) alongside the public lease, and CAKE
+  egress/ingress shaping via `wan-ifb`.
+
+The full fabric story — which switch port carries what, why VLAN 10 is trunked untranslated,
+and the multi-ONT plan — is in [sites/home/switches.md](sites/home/switches.md); the
+`my.homeRouter.*` options (`dns.wanSkipBroadcasts`, `firewall.untrustedRejectV4`) let each box
+tell `routing-common` about subnets sharing its WAN interface.
+
+## The AS211024 L2 mesh
+
+The edge routers are joined by a layer-2 mesh, defined once as `nixos.vpns.l2.as211024` in
+[`nixos/boxes/colony/vms/estuary/default.nix`](../nixos/boxes/colony/vms/estuary/default.nix)
+and realised on each member by the [`l2mesh` module](../nixos/modules/l2mesh.nix):
+
+- **Members**: `estuary`, `river`, `stream`, `britway`, peering on their public addresses.
+- **Transport**: VXLAN (VNI 211024, UDP port 4789) with static per-peer FDB entries and
+  UDP-encapsulated IPsec in transport mode via Libreswan (authentication-only by default;
+  `security.encrypt` would switch ESP from `null-sha256` to AES-GCM). The PSK is the shared
+  `l2mesh/as211024.key` secret, expanded into `/run/l2mesh.secrets` at `ipsec` start.
+- **Overlay addressing**: `10.100.50.0/24` / `2a0e:97c0:4df::/64`; the interface MTU is
+  computed from the physical MTU minus VXLAN/UDP/IPsec overhead. Each router holds
+  `10.100.50.<n>` (estuary `.1`, river `.2`, stream `.3`, britway `.5`).
+- **Routing over it**: the home routers route the colony prefixes via estuary and the
+  Tailscale prefixes via britway; estuary and britway route the home prefixes via the
+  `10.100.50.4` VIP. The home IPv6 **default** route also runs over the mesh, via britway
+  (hence the recursor's IPv4-only upstream pinning noted in
+  [`routing-common/dns.nix`](../nixos/boxes/home/routing-common/dns.nix)). The `nftTrust`
+  snippet in `lib.my.c.as211024` lets trusted inter-site traffic (colony, home, mesh,
+  Tailscale prefixes) through the edge firewalls.
+
+## BGP
+
+Both edge routers run `bird2` as AS211024:
+
+- **estuary** ([`bgp.nix`](../nixos/boxes/colony/vms/estuary/bgp.nix)): full table from
+  ColoClue (AS8283) over two sessions per family, plus IPv6 transit from iFog and Hurricane
+  Electric; peering at the Frys-IX, NL-ix and FogIXP route servers and with various networks
+  (Meta, Cloudflare, Apple, LUJE.net, …); a `bgp.tools` monitoring session. Static routes pull
+  the customer VIP blocks out of the `base` network and the internal/home IPv6 prefixes out of
+  the mesh.
+- **britway** ([`bgp.nix`](../nixos/boxes/britway/bgp.nix)): Vultr transit (AS64515,
+  MD5-passworded from a secret) and a `bgp.tools` session, originating the internal, colony
+  and home IPv6 prefixes.
+
+## WireGuard point-to-point tunnels
+
+Separate from the mesh, estuary terminates several networkd-managed WireGuard tunnels (private
+keys in per-box secrets):
+
+| Tunnel | Port | Remote / prefix | Notes |
+|---|---|---|---|
+| estuary ↔ kelder | 51820 | kelder holds `94.142.242.254/32` | kelder's public presence is a colony /32 routed over the tunnel |
+| estuary ↔ hillcrest | 51822 | `10.100.5.0/30` | point-to-point /30 out of `p2pTunnels` |
+| estuary ↔ john-valorant | 51823 | `10.100.5.4/30` | point-to-point /30 out of `p2pTunnels` |
+
+Additionally, the `qclk` container on `shill` runs its own WireGuard instances on port 51821
+out of `10.100.4.0/24`, and `britnet` hosts a road-warrior style WireGuard VPN on port 51820
+serving `10.200.0.0/24` / `fdfb:5ebf:6e84::/64`.
+
+## Tailscale / headscale
+
+Tailscale runs against a self-hosted **headscale** control plane on britway
+([`tailscale.nix`](../nixos/boxes/britway/tailscale.nix)) at `https://hs.nul.ie`: Google OIDC
+login, SQLite state, MagicDNS under `ts.nul.ie`, and split DNS that resolves the colony and
+home internal domains through their site resolvers. The tailnet prefixes are
+`100.64.0.0/10` / `fd7a:115c:a1e0::/48`.
+
+Notable nodes:
+
+- **waffletail** (container on `shill`) — the colony subnet router: advertises the colony
+  prefixes, acts as an exit node, and SNATs tailnet traffic into the colony networks.
+- **britway** — advertises the home prefixes (routed via the mesh) and is also an exit node.
+- Other boxes join with the shared `tailscale-auth.key` auth-key secret.
@@ -0,0 +1,14 @@
+# Remote boxes
+
+The "remote" group covers the boxes that live outside the `colony` and `home` sites: the two
+edge VPSes (`britway` in London, `britnet` in Birmingham) and the `kelder` site — a secondary
+server at a remote location, linked back to colony over WireGuard and acting as a NixOS
+container host.
+
+| Box | What it is | Docs |
+| --- | --- | --- |
+| `britway` | Vultr VPS (London, `lon1`): Headscale control plane, Tailscale exit node, BGP edge, nginx | [britway.md](britway.md) |
+| `britnet` | VPS (Birmingham, `bhx1`): Tailscale exit node / WireGuard hub | [britnet.md](britnet.md) |
+| `kelder` | Secondary home server (`hentai.engineer`): container host, Samba, DDNS | [kelder.md](kelder.md) |
+| `kelder-acquisition` | Media stack container on `kelder` (Transmission over VPN, *arr, Jellyfin) | [kelder-acquisition.md](kelder-acquisition.md) |
+| `kelder-spoder` | Web container on `kelder` (Nextcloud + nginx reverse proxy) | [kelder-spoder.md](kelder-spoder.md) |
@@ -0,0 +1,39 @@
+# britnet
+
+A small VPS in Birmingham (`bhx1`) acting as a second Tailscale/WireGuard egress point — a
+narrower gateway role than [`britway`](britway.md) (no control plane, no BGP).
+
+- **Source:** [`nixos/boxes/britnet.nix`](../../nixos/boxes/britnet.nix)
+- **Host:** VPS (Birmingham, `bhx1`; provider uplink assignment `allhost`)
+
+## Role
+
+- **Tailscale exit node** — logs into the headscale on [`britway`](britway.md)
+  (`--login-server=https://hs.nul.ie`) with `--advertise-exit-node`.
+- **WireGuard hub** — `wg0` listens on UDP 51820 on the `vpn` network
+  (`10.200.0.0/24` / `fdfb:5ebf:6e84::/64`), with a single peer at `10.200.0.10` /
+  `fdfb:5ebf:6e84::10`.
+- **NAT gateway** — traffic arriving on `tailscale0`/`wg0` is forwarded out `veth0` and SNATed
+  to the `allhost` v4/v6 addresses.
+
+## Network assignments
+
+<!-- assignments: britnet -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| britnet | allhost | `77.74.199.67/24 gw 77.74.199.1` | `2a12:ab46:5344:99::a/64 gw 2a12:ab46:5344::1` | bhx1.int.nul.ie |  |
+| britnet | vpn | `10.200.0.1/24` | `fdfb:5ebf:6e84::1/64` | — |  |
+<!-- assignments-end -->
+
+## Networking
+
+- The provider interface is renamed to `veth0` by MAC. The v6 default gateway
+  (`2a12:ab46:5344::1`) sits off-subnet, so a link-scope route is added to reach it.
+- `wg0` is a networkd WireGuard netdev keyed from `britnet/wg.key`; RA is disabled on it.
+- Upstream DNS is hardcoded to Cloudflare (`1.1.1.1` / `1.0.0.1`).
+- `iperf3` runs with an open port for bandwidth testing.
+
+## Notable config files
+
+- [`nixos/boxes/britnet.nix`](../../nixos/boxes/britnet.nix) — the whole box (single file).
@@ -0,0 +1,51 @@
+# britway
+
+A Vultr VPS in London (`lon1`) acting as the network edge node: the self-hosted Tailscale
+control plane, a tailnet exit node, and the BGP speaker for AS211024.
+
+- **Source:** [`nixos/boxes/britway/`](../../nixos/boxes/britway)
+- **Host:** VPS at Vultr (London, `lon1`)
+
+## Role
+
+- **Headscale** — the self-hosted Tailscale control plane at `hs.nul.ie`; every other box's
+  `tailscaled` logs in here (`--login-server=https://hs.nul.ie`). Google OIDC for auth,
+  SQLite state, MagicDNS under `ts.nul.ie`, and split DNS pointing the colony/home domains
+  at their internal resolvers.
+- **Tailscale exit node** — advertises `--advertise-exit-node` plus routes to the home v4/v6
+  prefixes; tailnet traffic is SNATed out `veth0` (v4 to the Vultr public IP, v6 to the
+  `as211024` mesh address).
+- **BGP edge** — `bird2` speaks BGP as AS211024 to Vultr transit (AS64515, separate v4/v6
+  sessions authenticated with a password from `britway/bgp-password-vultr.conf`) and exports
+  everything to a `bgp.tools` monitoring session. It originates the internal, colony and home
+  v6 prefixes (`2a0e:97c0:4df::/48`, `2a0e:97c0:4d2::/48`, `2a0e:97c0:4d0::/48`).
+- **nginx** — reverse proxy fronting headscale (`hs.nul.ie` → `localhost` headscale port),
+  with a wildcard ACME cert for `nul.ie` issued via Cloudflare DNS.
+
+## Network assignments
+
+<!-- assignments: britway -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| britway | as211024 | `10.100.50.5/24` | `2a0e:97c0:4df:0:2::1/64` | — |  |
+| britway | vultr | `45.76.141.188/23 gw 45.76.140.1` | `2001:19f0:7402:128b::1/64` | lon1.int.nul.ie |  |
+<!-- assignments-end -->
+
+## Networking
+
+- Two assignments: `vultr` on the provider interface `veth0` (renamed by MAC), and `as211024`
+  on the l2mesh VXLAN interface (`my.vpns.l2`, a mesh shared with `estuary` and the home
+  routers `river`/`stream`).
+- Static routes steer colony/home v4 traffic over the `as211024` mesh. A separate `ts-extra`
+  routing table (1337) with a policy rule on `tailscale0` ingress sends Tailscale-sourced v6
+  traffic for colony via `estuary`, while the box's own v6 uses WAN.
+- The firewall trusts the `as211024` prefixes (`lib.my.c.as211024.nftTrust`) and
+  `tailscale0`; `iperf3` runs with an open port for bandwidth testing.
+
+## Notable config files
+
+- [`nixos/boxes/britway/default.nix`](../../nixos/boxes/britway/default.nix) — system, assignments, networkd, firewall/SNAT.
+- [`nixos/boxes/britway/bgp.nix`](../../nixos/boxes/britway/bgp.nix) — `bird2` config (Vultr transit, `bgp.tools`).
+- [`nixos/boxes/britway/nginx.nix`](../../nixos/boxes/britway/nginx.nix) — nginx vhosts + ACME.
+- [`nixos/boxes/britway/tailscale.nix`](../../nixos/boxes/britway/tailscale.nix) — headscale + the tailnet node itself.
@@ -0,0 +1,48 @@
+# kelder-acquisition
+
+The media acquisition stack for the kelder site — Transmission (forced over VPN), the *arrs
+and Jellyfin in one NixOS container.
+
+- **Source:** [`nixos/boxes/kelder/containers/acquisition/`](../../nixos/boxes/kelder/containers/acquisition)
+- **Host:** NixOS container on [`kelder`](kelder.md)
+
+## Role
+
+- **Transmission** (`transmission_4`) — BitTorrent client bound to the VPN
+  (`bindsTo systemd-networkd-wait-online@vpn.service`); peer port 26180 (forwarded in the
+  AirVPN config), 20 MiB/s down / 1 MiB/s up limits, ratio limit 2.0. Downloads land in
+  `/mnt/media/downloads/torrents` (`/mnt/media` is bind-mounted from the host's
+  `/mnt/storage/media`).
+- **Jackett, Radarr, Sonarr** — indexer + media managers, in the shared `media` group with
+  `UMask 0002`.
+- **Jellyfin** — streaming with Intel hardware transcoding: `/dev/dri` is bind-mounted from
+  the host, `intel-vaapi-driver`/`intel-ocl` are installed and the `jellyfin` user is in the
+  `render` group.
+
+## Network assignments
+
+<!-- assignments: kelder-acquisition -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| acquisition-ctr | internal | `172.16.64.2/24 gw 172.16.64.1` | — | hentai.engineer |  |
+<!-- assignments-end -->
+
+## Networking
+
+- `internal` assignment (name `acquisition-ctr`) on the host's `ctrs` bridge, MTU 1460 to
+  match the site WAN.
+- All non-site traffic goes over an AirVPN WireGuard tunnel (`vpn` netdev, MTU 1320, AirVPN
+  IE endpoint) using fwmark 42 / route table 51820: policy rules keep traffic to and from the
+  kelder prefixes on the main table and push everything else via the VPN.
+- An nftables input chain drops new TCP connections from the VPN interface except the
+  Transmission peer port; the web UI ports (9091 Transmission, 9117 Jackett, 7878 Radarr,
+  8989 Sonarr, 8096 Jellyfin) are accepted from the site. When built as a dev VM, those ports
+  are forwarded to the host.
+- Sonarr still needs the EOL .NET 6 runtime, allowed via
+  `nixpkgs.config.permittedInsecurePackages`.
+
+## Notable config files
+
+- [`nixos/boxes/kelder/containers/acquisition/default.nix`](../../nixos/boxes/kelder/containers/acquisition/default.nix) — services and users.
+- [`nixos/boxes/kelder/containers/acquisition/networking.nix`](../../nixos/boxes/kelder/containers/acquisition/networking.nix) — AirVPN tunnel + firewall.
@@ -0,0 +1,54 @@
+# kelder-spoder
+
+The web container for the kelder site: Nextcloud plus an nginx (OpenResty) reverse proxy for
+the site's services.
+
+- **Source:** [`nixos/boxes/kelder/containers/spoder/`](../../nixos/boxes/kelder/containers/spoder)
+- **Host:** NixOS container on [`kelder`](kelder.md)
+
+## Role
+
+- **Nextcloud** (`nextcloud32`) at `cloud.hentai.engineer` (trusted alias
+  `cloud-local.hentai.engineer`), SQLite backend, data in `/mnt/storage/nextcloud`
+  (`/mnt/storage` is bind-mounted from the host).
+- **nginx reverse proxy** (`openresty`) terminating TLS for the site's public vhosts, with a
+  wildcard ACME cert for `hentai.engineer` via Cloudflare DNS. The kelder host forwards
+  `http`/`https` to this container; uploads are unlimited (`clientMaxBodySize = 0`) for
+  Nextcloud's sake.
+
+## Network assignments
+
+<!-- assignments: kelder-spoder -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| spoder-ctr | internal | `172.16.64.3/24 gw 172.16.64.1` | — | hentai.engineer |  |
+<!-- assignments-end -->
+
+## Proxy vhosts
+
+All under `hentai.engineer`, each with a `*-local` alias:
+
+| vhost | Target | Auth |
+| --- | --- | --- |
+| `monitor` | `netdata` on the kelder host (:19999) | `htpasswd` |
+| `kontent` | Jellyfin on `kelder-acquisition` (:8096, incl. websocket) | — |
+| `torrents` | Transmission on `kelder-acquisition` (:9091) | `htpasswd` |
+| `jackett` | Jackett on `kelder-acquisition` (:9117) | `htpasswd` |
+| `radarr` | Radarr on `kelder-acquisition` (:7878) | `htpasswd` |
+| `sonarr` | Sonarr on `kelder-acquisition` (:8989) | `htpasswd` |
+| `cloud` | Nextcloud (local) | — |
+
+An `init_worker_by_lua` timer polls `v4.ident.me` every 5 minutes to track the site's public
+IP; it feeds `localRedirect` rewrites (bounce public-IP clients to the `*-local` name) that
+are currently **disabled** (commented out — Virgin Media filters DNS answers containing local
+IPs, so the split doesn't work as intended).
+
+## Networking
+
+- `internal` assignment (name `spoder-ctr`) on the host's `ctrs` bridge, MTU 1420.
+
+## Notable config files
+
+- [`nixos/boxes/kelder/containers/spoder/default.nix`](../../nixos/boxes/kelder/containers/spoder/default.nix) — Nextcloud + ACME.
+- [`nixos/boxes/kelder/containers/spoder/nginx.nix`](../../nixos/boxes/kelder/containers/spoder/nginx.nix) — reverse proxy vhosts.
@@ -0,0 +1,67 @@
+# kelder
+
+Secondary home server at a remote site, domain `hentai.engineer`. Linked back to colony over
+WireGuard and acting as a NixOS container host (like `shill`/`sfh`).
+
+- **Source:** [`nixos/boxes/kelder/`](../../nixos/boxes/kelder)
+- **Host:** physical (Intel; LTS kernel, `kvm-intel`, IOMMU on)
+
+## Role
+
+- **Container host** — runs two NixOS containers on the `ctrs` bridge
+  (`my.containers.instances`): `kelder-acquisition` and `kelder-spoder` (below).
+- **Public services via colony** — a WireGuard tunnel (`estuary` netdev) connects to colony's
+  `estuary` box, which DNATs public traffic to kelder's `/32`; policy routing (table 51820 +
+  connmark) sends replies back through the tunnel while ordinary traffic uses the LAN.
+  kelder's own NAT forwards `http`/`https` on to `kelder-spoder`.
+- **Nextcloud host** — served from the `kelder-spoder` container.
+- **Samba** — the `storage` share backed by `/mnt/storage`, with `nmbd` and `samba-wsdd` for
+  Windows discovery.
+- **DDNS** — a `ddns-update` timer (every 10 min) runs `dns_update.py` to sync the
+  `hentai.engineer` and `kelder-local.hentai.engineer` Cloudflare records with the address on
+  `et1g0`.
+
+## Network assignments
+
+<!-- assignments: kelder -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| kelder-ctrs | ctrs | `172.16.64.1/24` | — | hentai.engineer |  |
+| kelder | estuary | `94.142.242.254/32` | — | — |  |
+<!-- assignments-end -->
+
+## Containers
+
+| Container | Role | Docs |
+| --- | --- | --- |
+| `kelder-acquisition` | Media stack (Transmission over AirVPN, Jackett/Radarr/Sonarr, Jellyfin) | [kelder-acquisition.md](kelder-acquisition.md) |
+| `kelder-spoder` | Nextcloud + nginx reverse proxy | [kelder-spoder.md](kelder-spoder.md) |
+
+The containers are not deploy targets (`my.deploy.enable = false`); they're managed through
+the host.
+
+## Networking
+
+- LAN on `et1g0` (renamed by MAC) with DHCP and MTU 1460 (`lib.my.c.kelder.ipv4MTU`); the
+  kelder v4 prefixes are masqueraded out of it.
+- The `estuary` WireGuard peer (`AllowedIPs = 0.0.0.0/0`, keepalive 25s) is combined with
+  rules that keep LAN traffic on the main table and only route tunnel-marked/owned traffic
+  via table 51820.
+
+## Services
+
+- `netdata` (proxied as `monitor.hentai.engineer` by `kelder-spoder`), `smartd`, `fstrim`,
+  LVM thin provisioning.
+- `minecraft-server` is present but **disabled** (`enable = false`); the firewall still opens
+  25565 tcp/udp.
+- Primary user `kontent` (in the `storage`/`media` groups).
+- Sets `system.nixos.distroName = "KelderOS"`, a custom Plymouth theme and an `amogus-beep`
+  boot jingle ([`boot.nix`](../../nixos/boxes/kelder/boot.nix)).
+
+## Notable config files
+
+- [`nixos/boxes/kelder/default.nix`](../../nixos/boxes/kelder/default.nix) — system, assignments, tunnel, NAT, containers.
+- [`nixos/boxes/kelder/boot.nix`](../../nixos/boxes/kelder/boot.nix) — Plymouth theme + boot beep.
+- [`nixos/boxes/kelder/containers/`](../../nixos/boxes/kelder/containers) — the two container definitions.
+- [`nixos/boxes/kelder/dns_update.py`](../../nixos/boxes/kelder/dns_update.py) — Cloudflare DDNS script.
@@ -0,0 +1,90 @@
+# colony
+
+The hosted dedicated server in Amsterdam (`ams1`) and the public-facing half of
+the boxes: almost everything reachable from the internet lives here.
+
+- **Internal domain:** `ams1.int.nul.ie` (`lib.my.c.colony.domain`)
+- **Public domain:** `nul.ie` — public services are published as `*.nul.ie`
+- **Source:** [`nixos/boxes/colony/`](../../../nixos/boxes/colony)
+
+## Shape
+
+`colony` is the physical VM host. It runs the VMs below; `shill` is itself a
+NixOS container host where most applications run.
+
+```
+colony (physical VM host, ams1)
+├── estuary ── edge router: WAN, firewall/NAT, DNS, BGP (AS211024), WireGuard
+├── shill ──── NixOS container host ──┬── middleman    (reverse proxy, ACME, nginx-sso, librespeed)
+│                                     ├── vaultwarden  (password manager)
+│                                     ├── colony-psql  (shared PostgreSQL)
+│                                     ├── chatterbox   (Matrix Synapse + bridges)
+│                                     ├── jackflix     (media stack)
+│                                     ├── object       (MinIO, Harmonia Nix cache, Sharry, HedgeDoc, wastebin)
+│                                     ├── toot         (Bluesky PDS; Mastodon disabled)
+│                                     ├── waffletail   (Tailscale subnet router / exit node)
+│                                     ├── qclk         (WireGuard management appliance)
+│                                     ├── gam          (Terraria server)
+│                                     └── jam          (raw nspawn customer container)
+├── whale2 ─── podman/OCI host for game servers
+├── git ────── Gitea + Gitea Actions runner
+├── mail ───── Debian VM running mailcow (not NixOS)
+└── darts ──── third-party/customer VM (opaque, not NixOS)
+```
+
+## Networks
+
+All internal space is carved out of `10.100.0.0/16` and `2a0e:97c0:4d2:10::/60`
+(`lib.my.c.colony.prefixes`):
+
+| Network | IPv4 CIDR | IPv6 CIDR | Purpose |
+|---|---|---|---|
+| `base` | `10.100.0.0/24` | `2a0e:97c0:4d2:10::/64` | Base LAN shared by `colony` and `estuary` |
+| `vms` | `10.100.1.0/24` | `2a0e:97c0:4d2:11::/64` | VM network |
+| `ctrs` | `10.100.2.0/24` | `2a0e:97c0:4d2:12::/64` | `shill` container network |
+| `oci` | `10.100.3.0/24` | `2a0e:97c0:4d2:13::/64` | `whale2` podman network |
+
+Public addressing — the WAN /24 (`94.142.240.44`), the `vip*` ranges shared by
+the VMs, and the customer /32s for `mail` / `darts` — terminates on `estuary`;
+see [estuary.md](estuary.md).
+
+## Machines
+
+| Machine | Role | Page |
+|---|---|---|
+| `colony` | Physical VM host (AMD, KVM, LVM-thin, `borgthin` backups → rsync.net) | [colony.md](colony.md) |
+| `estuary` | Edge router: WAN, firewall/NAT, DNS, BGP (AS211024), WireGuard | [estuary.md](estuary.md) |
+| `shill` | NixOS container host (most applications) | [shill.md](shill.md) |
+| `whale2` | podman/OCI game-server host | [whale2.md](whale2.md) |
+| `git` | Gitea + Gitea Actions runner | [git.md](git.md) |
+| `mail` | Debian VM running mailcow (not NixOS) | [mail.md](mail.md) |
+| `darts` | Third-party/customer VM (not NixOS) | [darts.md](darts.md) |
+
+### `shill` containers
+
+Each has its own page under `shill/containers/`:
+
+| Container | Role | Page |
+|---|---|---|
+| `middleman` | Front-end nginx reverse proxy, ACME, nginx-sso, librespeed | [middleman](shill/containers/middleman.md) |
+| `vaultwarden` | Vaultwarden password manager | [vaultwarden](shill/containers/vaultwarden.md) |
+| `colony-psql` | Shared PostgreSQL for colony services | [colony-psql](shill/containers/colony-psql.md) |
+| `chatterbox` | Matrix homeserver + bridges | [chatterbox](shill/containers/chatterbox.md) |
+| `jackflix` | Media stack (Jellyfin, *arr, Transmission, PhotoPrism, copyparty) | [jackflix](shill/containers/jackflix.md) |
+| `object` | MinIO (S3), Harmonia Nix cache, Sharry, HedgeDoc, wastebin | [object](shill/containers/object.md) |
+| `toot` | Bluesky PDS (Mastodon disabled) | [toot](shill/containers/toot.md) |
+| `waffletail` | Tailscale subnet router / exit node | [waffletail](shill/containers/waffletail.md) |
+| `qclk` | WireGuard management appliance | [qclk](shill/containers/qclk.md) |
+| `gam` | Terraria server | [gam](shill/containers/gam.md) |
+
+## Non-NixOS VMs
+
+Two VMs are declared in `colony`'s `my.vms.instances` — so `colony` runs them
+and routes/firewalls their traffic — but they are **not** managed as NixOS
+systems by this repo:
+
+- **`mail`** — a Debian VM running [mailcow](https://mailcow.email/)
+  (`mail.nul.ie`). ACME certificates are pushed to it from `middleman`.
+  See [mail.md](mail.md).
+- **`darts`** — an opaque third-party/customer VM, given a routed public /32
+  and IPv6 /64 and otherwise left alone. See [darts.md](darts.md).
@@ -0,0 +1,86 @@
+# colony
+
+The physical dedicated server in Amsterdam (`ams1`) and the VM host for
+everything at the colony site.
+
+- **Source:** [`nixos/boxes/colony/default.nix`](../../../nixos/boxes/colony/default.nix)
+  (VM instances in [`nixos/boxes/colony/vms/default.nix`](../../../nixos/boxes/colony/vms/default.nix))
+- **Host:** bare metal (this *is* the physical box)
+- **nixpkgs:** `mine-stable`
+
+## Role
+
+Bare-metal AMD host. It does little application work itself — its job is to run
+the VMs and provide them with storage, networking and backups.
+
+- **Virtualisation:** QEMU/KVM (`kvm-amd`, IOMMU on) driven by the `my.vms`
+  module: each entry in `my.vms.instances` becomes a `vm@<name>` systemd
+  service running `qemu-kvm` with UEFI, a QMP/monitor socket under
+  `/run/vms/<name>/`, TAP networking and optional PCI passthrough
+  (`hostDevices`, bound to `vfio-pci`). `estuary` gets the WAN NIC this way.
+- **Storage:** LVM-thin (`services.lvm.boot.thin`) in the `main` VG; VM disks
+  are logical volumes (`vm-<name>-<disk>`, see the `lib.my.vm.disk` /
+  `lvmDisk` helpers). `/persist` holds host state, `/mnt/backup` the local
+  borg repo. Only the boot-critical LVs are activated in the initrd; the rest
+  come up via `lvm-activate-main.service`.
+- **Backups:** `my.borgthin` job `main` snapshots the persist/data LVs of the
+  host and its VMs into `/mnt/backup/main`; `borgthin-rsync.service` then
+  rsyncs the repo to rsync.net and `rsync-lvm-meta.service` ships the LVM
+  metadata alongside (both run idle-priority, after the borg job).
+- **Monitoring/health:** netdata (freeipmi, ignoring the VCCM sensor), `smartd`
+  (logging to `/var/log/smartd`), `rasdaemon`, `fstrim` at 04:45 (before the
+  05:00 backup).
+
+## Network assignments
+
+<!-- assignments: colony -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| colony (vm) | internal | `94.142.241.224/32` | `2a0e:97c0:4d2:10::2/64` | ams1.int.nul.ie |  |
+| colony-routing | routing | `10.100.0.2/24 gw 10.100.0.1` | — | ams1.int.nul.ie |  |
+| colony-vms | vms | `10.100.1.1/24` | `2a0e:97c0:4d2:11::1/64` | ams1.int.nul.ie |  |
+<!-- assignments-end -->
+
+## Networking
+
+- Two bridges: `base` (the colony base network, shared with `estuary`) and
+  `vms` (the VM network). Dummy interfaces (`base0`, `vms0`) keep the bridges
+  up in networkd's eyes so dependent VMs can start.
+- `colony` sends RAs on `vms` (DNS = `estuary`'s base address) and carries
+  static routes for the downstream prefixes: `ctrs` via `shill`, `oci` via
+  `whale2`, plus the Tailscale, `qclk` and `jam` prefixes via `shill`.
+- `estuary` is the default gateway (via the `base` bridge); `colony`'s own
+  public-facing address is its `internal` assignment (a `vip1` /32, alt name
+  `vm`).
+- The customer VMs attach to dedicated TAP devices (`vm-mail`, `vm-darts`)
+  which are **not** bridged: networkd puts the point-to-point /32
+  (`lib.my.c.colony.custRouting`) and the customer's IPv6 /64 on each, sends
+  RAs, and link-routes the customer's public /32 down the tap.
+- `my.firewall` trusts the `vms` bridge, DNATs the shared
+  `lib.my.c.colony.firewallForwards` list for traffic addressed to `estuary`'s
+  public IP (so the port forwards also work from inside), and forwards the
+  customer prefixes through with minimal filtering ("trust for now").
+
+## VMs
+
+Declared in `my.vms.instances` (`cpus`/`threads` are QEMU `smp` values):
+
+| VM | Cores | Threads | Memory | MAC | Disks |
+|---|---|---|---|---|---|
+| `estuary` | 2 | 2 | 3 GiB | `52:54:00:15:1a:53` (`base`) | `esp` / `nix` / `persist` LVs + WAN NIC passthrough |
+| `shill` | 12 | 2 | 40 GiB | `52:54:00:27:3d:5c` (`vms`) | `esp` / `nix` / `persist` + `media` / `minio` / `nix-cache` / `jam` LVs |
+| `whale2` | 8 | 2 | 16 GiB | `52:54:00:d5:d9:c6` (`vms`) | `esp` / `nix` / `persist` + `oci` LV |
+| `git` | 12 | 2 | 40 GiB | `52:54:00:75:78:a8` (`vms`) | `esp` / `nix` / `persist` / `oci` + `git` / `gitea-actions-cache` LVs |
+| `mail` | 3 | 2 | 6 GiB | `52:54:00:a8:d1:03` (`vm-mail` tap) | `root` / `data` LVs |
+| `darts` | 4 | 2 | 16 GiB | `52:54:00:a8:29:cd` (`vm-darts` tap) | `root` LV + `darts-media` / `darts-ext` LVs |
+
+`estuary`, `shill`, `whale2` and `git` are NixOS systems with their own pages
+(see the [README](README.md#machines)); `mail` and `darts` are not (see
+[Non-NixOS VMs](README.md#non-nixos-vms)).
+
+## Notable config files
+
+- [`nixos/boxes/colony/default.nix`](../../../nixos/boxes/colony/default.nix) — host hardware, networkd, firewall, backups.
+- [`nixos/boxes/colony/vms/default.nix`](../../../nixos/boxes/colony/vms/default.nix) — `my.vms.instances` for all six VMs.
+- [`nixos/modules/vms.nix`](../../../nixos/modules/vms.nix) — the `my.vms` module itself.
@@ -0,0 +1,27 @@
+# darts
+
+An opaque third-party/customer VM. Declared in `colony`'s `my.vms.instances`
+but **not a NixOS system**: this repo knows nothing about what runs inside it
+and doesn't manage it.
+
+- **Source (host-side only):** the `darts` instance in
+  [`nixos/boxes/colony/vms/default.nix`](../../../nixos/boxes/colony/vms/default.nix)
+  and the `90-vm-darts` network in
+  [`nixos/boxes/colony/default.nix`](../../../nixos/boxes/colony/default.nix)
+- **Host:** VM on `colony`
+
+## Role
+
+- Customer/dedicated VM, left alone beyond hosting and connectivity.
+- 4 cores, 16 GiB RAM; a `root` LV plus `darts-media` and `darts-ext` LVs from
+  the `media`/`ext` volume groups.
+
+## Networking
+
+- Same customer-VM pattern as [`mail`](mail.md): dedicated unbridged TAP
+  (`vm-darts`, MAC `52:54:00:a8:29:cd`), point-to-point address
+  (`custRouting.darts-vm`) on the host side, link-routed public /32
+  `94.142.242.255`, and the IPv6 /64 `2a0e:97c0:4d2:2001::/64` with RAs.
+- DNS: `darts-cust.ams1.int.nul.ie`. Like the other customer prefixes, its
+  inbound traffic is accepted by `estuary` without per-port filtering and
+  forwarded on by `colony`.
@@ -0,0 +1,145 @@
+# estuary
+
+The colony edge router and firewall — the box that holds colony's public IPs
+and connects everything else at the site to the internet.
+
+- **Source:** [`nixos/boxes/colony/vms/estuary/`](../../../nixos/boxes/colony/vms/estuary)
+  (`default.nix`, `bgp.nix`, `dns.nix`, `bandwidth.nix`)
+- **Host:** VM on `colony` (gets the WAN NIC by PCI passthrough)
+- **nixpkgs:** `mine`
+
+## Role
+
+- **Edge routing / firewall / NAT:** owns the colony public IPv4/IPv6
+  (`94.142.240.44/24`, `2a02:898:0:20::329:1/64`), NATs outbound traffic, and
+  port-forwards inbound services (`my.firewall.nat.forwardPorts` driven by the
+  shared `lib.my.c.colony.firewallForwards` list): HTTP/S and Matrix
+  federation to `middleman`, git to `git`, game ports to the OCI servers on
+  `whale2` and to `gam`, Tailscale to `waffletail`, and the `qclk` WireGuard
+  port.
+- **DNS:** PowerDNS authoritative server *and* recursor (see below).
+- **BGP:** BIRD2 speaking AS211024 with upstreams, IXP route servers and
+  direct peers (see below).
+- **VPNs:** member of the `as211024` L2 VXLAN mesh and endpoint for three
+  point-to-point WireGuard tunnels (see below).
+- **Misc:** `iperf3` server, netdata.
+
+## Network assignments
+
+<!-- assignments: estuary -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| estuary | as211024 | `10.100.50.1/24` | `2a0e:97c0:4df::1/64` | — |  |
+| estuary-vm-base | base | `10.100.0.1/24` | `2a0e:97c0:4d2:10::1/64` | ams1.int.nul.ie |  |
+| estuary-vm (fw) | internal | `94.142.240.44/24 gw 94.142.240.254` | `2a02:898:0:20::329:1/64 gw 2a02:898:0:20::1` | ams1.int.nul.ie |  |
+<!-- assignments-end -->
+
+## WAN and IXP VLANs
+
+- `wan` — the passed-through `igb` NIC (9000 MTU, enlarged rings). It carries
+  the plain upstream uplink (static v4/v6 with gateways from the `internal`
+  assignment) plus the tagged `ifog` VLAN.
+- `ifog` (VLAN 409) is an iFog QinQ transport that carries the IXP VLANs as
+  nested tags:
+
+| Interface | VLAN | IPv4 | IPv6 | Purpose |
+|---|---|---|---|---|
+| `frys-ix` | 701 | `185.1.160.196/23` | `2001:7f8:10f::3:3850:196/64` | Frys-IX peering LAN |
+| `nl-ix` | 1845 | `193.239.116.145/22` | `2001:7f8:13::a521:1024:1/64` | NL-ix peering LAN |
+| `fogixp` | 1147 | `185.1.147.159/24` | `2001:7f8:ca:1::159/64` | FogIXP peering LAN |
+| `ifog-transit` | 702 | — | `2a0c:9a40:100f:370::2/64` | iFog IPv6 transit |
+
+The IXP interfaces run at 1500 MTU with DHCP/RA/LLDP off; an nftables `ixp`
+chain rejects non-IP/ARP ethertypes in both directions.
+
+- `base` — colony base network; sends RAs and serves DNS to the site, and
+  routes the `vms`/`ctrs`/`oci`, Tailscale, `qclk`, `vip*` and customer
+  prefixes back via `colony`.
+- `as211024` — the L2 mesh interface (see VPNs).
+
+## Firewall and NAT
+
+`my.firewall` (nftables). Besides the port forwards, `extraRules` defines:
+
+- `routing-tcp` / `routing-udp` chains — the inbound allow-list for new
+  connections from `wan`/`as211024`/IXPs towards internal services (SSH
+  anywhere, otherwise per-service v4/v6 rules mirroring `firewallForwards`).
+- `filter-routing` — applied to `wan`/`as211024`/IXPs → `base` forwards;
+  customer prefixes (`mail`/`darts` v4, `cust.v6`) are accepted wholesale, the
+  rest goes through the `routing-*` chains.
+- SNAT: everything from `prefixes.all.v4` leaving non-`as211024` interfaces is
+  NATed to the public IP; the WireGuard tunnel prefixes get their own SNAT
+  addresses.
+- DNS redirect: DNS traffic arriving at estuary's own public addresses is
+  redirected to port 5353 (the authoritative server) — see below.
+
+## DNS
+
+Both halves are PowerDNS ([`dns.nix`](../../../nixos/boxes/colony/vms/estuary/dns.nix)):
+
+- **Authoritative** (`my.pdns.auth`) listens on `0.0.0.0:5353` / `[::]:5353` as
+  primary for `ams1.int.nul.ie`, `100.10.in-addr.arpa` and the
+  `2a0e:97c0:4d2::/48` reverse zone. Zone contents are largely generated from
+  `allAssignments` (`lib.my.dns.fwdRecords` / `ptrRecords` / `ptr6Records`);
+  `ALIAS` records (with `expand-alias`) point the zone apex at estuary itself.
+  AXFR is allowed to HE.net's secondary (`216.218.133.2` / `2001:470:600::2`),
+  and `_acme-challenge` is a LUA `TXT` record answered from a file (used for
+  DNS-01 issuance). Public DNS reaches it via the NAT redirect of port 53 to
+  5353; the `base` side also accepts DNS directly.
+- **Recursor** (`my.pdns.recursor`, `pdns-recursor`) listens on localhost and
+  the `base` addresses, serving `prefixes.all` and the Tailscale prefixes. The
+  authoritative zones are forwarded back to `127.0.0.1:5353` (with NOTIFY
+  support so changes show up immediately), and a small Lua `preresolve` hook
+  rewrites `nix-cache.nul.ie` to `http.ams1.int.nul.ie` so cache traffic stays
+  on-site.
+
+## BGP
+
+BIRD2 ([`bgp.nix`](../../../nixos/boxes/colony/vms/estuary/bgp.nix)) speaking **AS211024**:
+
+- **Upstreams:** ColoClue (AS8283, `euNetworks` 2/3, v4+v6); iFog IPv6 transit
+  (AS34927); Hurricane Electric IPv6 over Frys-IX (AS6939).
+- **IXP route servers:** Frys-IX (AS56393), NL-ix (AS34307, depref'd by 1),
+  FogIXP (AS47498).
+- **Direct peers:** LUJE.net (AS212855, on ColoClue/Frys-IX/FogIXP + multihop
+  labs), jurrian (AS212635), Meta (AS32934, Frys-IX/NL-ix), Cloudflare
+  (AS13335, Frys-IX), Apple (AS714, NL-ix), HE (AS6939, Frys-IX v4).
+- **Monitoring:** a multihop session to the bgp.tools collector (AS212232)
+  exporting everything with `add paths`.
+- Statics originate the site prefixes: the `vip*` ranges via `base`, the
+  ams1/internal v6 prefixes via `base`/`as211024`, and the home /48 towards
+  the home routers' `as211024` VIP.
+
+Currently disabled (commented out): efero transit over FogIXP ("not working so
+well lately") and the NL-ix Cloudflare sessions.
+
+## VPNs
+
+- **`as211024` L2 mesh** (`my.vpns.l2`, the `l2mesh` module): VXLAN (VNI
+  211024, UDP-encapsulated) secured with libreswan IPsec, meshing estuary with
+  the home routers `river`/`stream` and `britway`. This carries the AS211024
+  anycast-ish internal address space between sites.
+- **WireGuard endpoints** (networkd `wireguard` netdevs, keys from agenix):
+  - `kelder` — tunnel to the remote `kelder` site, port `51820`.
+  - `hillcrest` — port `51822`, point-to-point /32 pair out of
+    `prefixes.hillcrest.v4`.
+  - `john-valorant` — port `51823`, same pattern out of
+    `prefixes.john-valorant.v4`.
+
+## Bandwidth management
+
+[`bandwidth.nix`](../../../nixos/boxes/colony/vms/estuary/bandwidth.nix)
+implements a ~95% WAN shaper: a 245 Mbit token-bucket filter on `wan`
+(outbound) and on an `ifb-wan` IFB device that ingress traffic is mirrored
+into (inbound), with [`bandwidth.py`](../../../nixos/boxes/colony/vms/estuary/bandwidth.py)
+as a `bandwidth-limiter` service that watches/utilises the link and can adjust
+the rate. **Currently disabled** — the file is not in estuary's `imports`
+(only `dns.nix` and `bgp.nix` are), so no shaping is applied.
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/estuary/default.nix`](../../../nixos/boxes/colony/vms/estuary/default.nix) — system, networkd, firewall, WireGuard, mesh membership.
+- [`nixos/boxes/colony/vms/estuary/bgp.nix`](../../../nixos/boxes/colony/vms/estuary/bgp.nix) — BIRD2 config.
+- [`nixos/boxes/colony/vms/estuary/dns.nix`](../../../nixos/boxes/colony/vms/estuary/dns.nix) — PowerDNS auth + recursor.
+- [`nixos/boxes/colony/vms/estuary/bandwidth.nix`](../../../nixos/boxes/colony/vms/estuary/bandwidth.nix) — WAN shaper (disabled, not imported).
@@ -0,0 +1,57 @@
+# git
+
+The Gitea VM — source hosting and CI for the boxes (`git.nul.ie`).
+
+- **Source:** [`nixos/boxes/colony/vms/git/`](../../../nixos/boxes/colony/vms/git)
+  (`default.nix`, `gitea.nix`, `gitea-actions.nix`)
+- **Host:** VM on `colony`
+- **nixpkgs:** `mine`
+
+## Role
+
+- **Gitea** ([`gitea.nix`](../../../nixos/boxes/colony/vms/git/gitea.nix)) — the
+  Git forge at `git.nul.ie` (self-registration disabled). Backed by PostgreSQL
+  on `colony-psql` (waiting on it via `lib.my.systemdAwaitPostgres`), LFS
+  enabled, with all object storage (incl. LFS and packages) on MinIO at
+  `s3.nul.ie` (bucket `gitea`, on `object`; the secret is spliced into
+  `app.ini` at startup). Mail goes out via `mail.nul.ie`, including the
+  issue-reply incoming-mail poller.
+- **Gitea Actions runner**
+  ([`gitea-actions.nix`](../../../nixos/boxes/colony/vms/git/gitea-actions.nix)) —
+  one Docker-mode instance (`main-docker`) on podman (privileged, `podman`
+  network), with labels for `node:24-trixie` and the self-built
+  `git.nul.ie/dev/actions-ubuntu:26.04` images. Runs as a fixed
+  `gitea-runner` user (not `DynamicUser`) so it can read its token; 8 h job
+  timeout; the action cache lives on a dedicated disk
+  (`/var/cache/gitea-runner`). This runner executes the repo's own
+  `.gitea/workflows/ci.yaml`.
+- **nginx** — terminates TLS for `git.nul.ie` (and a default vhost) and
+  proxies to Gitea on `:3000`. ACME (Let's Encrypt, production) issues
+  `nul.ie` + `*.nul.ie` via the Cloudflare DNS-01 challenge.
+- **podman** — local container backend for the runner;
+  `/var/lib/containers` is an XFS data disk, and the default
+  `10.88.0.0/16` podman subnet is allowed to forward.
+
+## Network assignments
+
+<!-- assignments: git -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| git-vm | internal | `94.142.241.117/32` | `2a0e:97c0:4d2:11::4/64` | ams1.int.nul.ie |  |
+| git-vm-routing | routing | `10.100.1.4/24 gw 10.100.1.1` | — | ams1.int.nul.ie |  |
+<!-- assignments-end -->
+
+## Storage
+
+- `/var/lib/gitea` — the `git` LV (repositories, config).
+- `/var/cache/gitea-runner` — the `gitea-actions-cache` LV.
+- `/var/lib/containers` — the `oci` LV (XFS with project quotas). Despite the
+  name this is local to the `git` VM and unrelated to `whale2`'s `oci`
+  network.
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/git/default.nix`](../../../nixos/boxes/colony/vms/git/default.nix) — VM config, nginx + ACME, podman, firewall.
+- [`nixos/boxes/colony/vms/git/gitea.nix`](../../../nixos/boxes/colony/vms/git/gitea.nix) — Gitea itself.
+- [`nixos/boxes/colony/vms/git/gitea-actions.nix`](../../../nixos/boxes/colony/vms/git/gitea-actions.nix) — the Actions runner.
@@ -0,0 +1,36 @@
+# mail
+
+A Debian VM running [mailcow](https://mailcow.email/) (`mail.nul.ie`) — the
+mail server for `nul.ie`. Declared in `colony`'s `my.vms.instances` but **not
+a NixOS system**: everything inside the VM is configured out of band.
+
+- **Source (host-side only):** the `mail` instance in
+  [`nixos/boxes/colony/vms/default.nix`](../../../nixos/boxes/colony/vms/default.nix)
+  and the `90-vm-mail` network in
+  [`nixos/boxes/colony/default.nix`](../../../nixos/boxes/colony/default.nix)
+- **Host:** VM on `colony`
+
+## Role
+
+- Runs the full mailcow stack (Postfix/Dovecot/SOGo/Rspamd) for `nul.ie`.
+  Other colony services send through it as `mail.nul.ie` (e.g. Gitea, and the
+  disabled Mastodon config).
+- 3 cores, 6 GiB RAM; `root` and `data` LVM disks (the `vm-mail-data` LV is
+  included in `colony`'s `borgthin` backups).
+
+## Networking
+
+- Attaches to a dedicated, unbridged TAP (`vm-mail`, MAC
+  `52:54:00:a8:d1:03`). `colony` puts the point-to-point address
+  (`custRouting.mail-vm`) on the host side, link-routes the public /32
+  `94.142.241.227` down the tap, and serves the IPv6 /64
+  `2a0e:97c0:4d2:2000::/64` with RAs.
+- DNS: `mail-vm.ams1.int.nul.ie` (and `mail.nul.ie` publicly, incl. the PTR in
+  estuary's reverse zone). `estuary` accepts traffic to the customer prefixes
+  without per-port filtering; `colony` forwards it on ("trust for now").
+
+## Notes
+
+- ACME certificates are issued on `middleman` and pushed to the VM over SSH
+  (`acme@mail.nul.ie mailcow-ssl-reload`, key `middleman/mailcow-ssh.key`);
+  the VM's SSH host key is pinned at `.keys/mail-vm-host.pub`.
@@ -0,0 +1,83 @@
+# shill
+
+The colony NixOS container host — most colony applications run as
+`systemd-nspawn` containers on this VM.
+
+- **Source:** [`nixos/boxes/colony/vms/shill/`](../../../nixos/boxes/colony/vms/shill)
+  (`default.nix`, `containers-ext.nix`, `containers/`)
+- **Host:** VM on `colony` (large: 12 cores, 40 GiB RAM)
+- **nixpkgs:** `mine`
+
+## Role
+
+- Runs the colony containers via `my.containers.instances`, each attached to
+  the `ctrs` bridge with its own address. The containers are full NixOS
+  systems rendered via `my.asContainer` and deployed as container profiles on
+  `shill` (`my.deploy.enable = false` — they are not standalone deploy
+  targets); the `my.containers` module wires up the nspawn units, `/persist`
+  bind mounts and store binds.
+- Provides shared data volumes to containers via bind mounts from LVM-backed
+  disks: `/mnt/media` (→ `middleman` read-only, `jackflix` read-write),
+  `/mnt/minio` and `/mnt/nix-cache` (→ `object`, both read-write).
+- Routes between the `vms` network and the `ctrs` container network: sends RAs
+  on `ctrs` (DNS = `estuary`'s base address) and routes the Tailscale prefixes
+  via `waffletail` and the `qclk` prefix via `qclk`. Applies the shared
+  `firewallForwards` DNAT for traffic addressed to `estuary`'s public IP, with
+  an nftables `ct mark 0x1337` SNAT hack so forwarded return traffic stays
+  symmetric.
+- Tuned for high connection counts (larger conntrack table, wider ephemeral
+  port range); netdata on 19999.
+
+## Network assignments
+
+<!-- assignments: shill -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| shill-vm-ctrs | ctrs | `10.100.2.1/24` | `2a0e:97c0:4d2:12::1/64` | ams1.int.nul.ie |  |
+| shill-vm (ctr) | internal | `94.142.241.225/32` | `2a0e:97c0:4d2:11::2/64` | ams1.int.nul.ie |  |
+| shill-vm-routing | routing | `10.100.1.2/24 gw 10.100.1.1` | — | ams1.int.nul.ie |  |
+<!-- assignments-end -->
+
+## Containers
+
+Defined under
+[`shill/containers/`](../../../nixos/boxes/colony/vms/shill/containers) and
+wired up in `shill`'s `my.containers.instances`. Each has its own page:
+
+| Container | IPv4 | IPv6 | Role | Page |
+|---|---|---|---|---|
+| `middleman` | `10.100.2.2` | `2a0e:97c0:4d2:12::2` | Reverse proxy, ACME, nginx-sso, librespeed | [middleman](shill/containers/middleman.md) |
+| `vaultwarden` | `10.100.2.3` | `2a0e:97c0:4d2:12::3` | Password manager | [vaultwarden](shill/containers/vaultwarden.md) |
+| `colony-psql` | `10.100.2.4` | `2a0e:97c0:4d2:12::4` | Shared PostgreSQL (14) | [colony-psql](shill/containers/colony-psql.md) |
+| `chatterbox` | `10.100.2.5` | `2a0e:97c0:4d2:12::5` | Matrix Synapse + bridges | [chatterbox](shill/containers/chatterbox.md) |
+| `jackflix` | `10.100.2.6` | `2a0e:97c0:4d2:12::6` | Media stack | [jackflix](shill/containers/jackflix.md) |
+| `object` | `10.100.2.7` | `2a0e:97c0:4d2:12::7` | MinIO, Harmonia Nix cache, Sharry, HedgeDoc, wastebin | [object](shill/containers/object.md) |
+| `toot` | `10.100.2.8` | `2a0e:97c0:4d2:12::8` | Bluesky PDS (Mastodon disabled) | [toot](shill/containers/toot.md) |
+| `waffletail` | `10.100.2.9` | `2a0e:97c0:4d2:12::9` | Tailscale subnet router / exit node | [waffletail](shill/containers/waffletail.md) |
+| `qclk` | `10.100.2.10` | `2a0e:97c0:4d2:12::a` | WireGuard management appliance | [qclk](shill/containers/qclk.md) |
+| `gam` | `10.100.2.11` | `2a0e:97c0:4d2:12::b` | Terraria server | [gam](shill/containers/gam.md) |
+
+### `jam`
+
+A one-off: [`containers-ext.nix`](../../../nixos/boxes/colony/vms/shill/containers-ext.nix)
+runs a raw `systemd-nspawn` container (not a `my.containers` instance, not
+NixOS) with its root on the `jam` LV, private user namespaces and a `ve-jam`
+veth. It gets the `jam` customer prefix (`prefixes.jam`, `jam-cust` in DNS)
+and SSH is forwarded to it from `shill`'s public IP port 60022.
+
+## Notes
+
+- `nix.settings.substituters` is forced to just `https://cache.nixos.org` —
+  `shill` sits next to the S3 cache on `object`, so it doesn't use it.
+- [`hercules.nix`](../../../nixos/boxes/colony/vms/shill/hercules.nix)
+  (Hercules CI agent + the `nix-cache-gc` timer for the S3 binary cache)
+  exists but is **currently disabled**: the file is not imported by
+  `shill/default.nix`.
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/shill/default.nix`](../../../nixos/boxes/colony/vms/shill/default.nix) — VM config, networkd, firewall, `my.containers.instances`.
+- [`nixos/boxes/colony/vms/shill/containers/default.nix`](../../../nixos/boxes/colony/vms/shill/containers/default.nix) — container imports.
+- [`nixos/boxes/colony/vms/shill/containers-ext.nix`](../../../nixos/boxes/colony/vms/shill/containers-ext.nix) — the `jam` nspawn container.
+- [`nixos/modules/containers.nix`](../../../nixos/modules/containers.nix) — the `my.containers` module.
@@ -0,0 +1,45 @@
+# chatterbox
+
+The Matrix homeserver for `nul.ie` (Synapse) and its bridges to other chat networks.
+[middleman](middleman.md) fronts it as `matrix.nul.ie` for clients and on `:8448` for
+federation.
+
+- **Source:** [`shill/containers/chatterbox.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/chatterbox.nix)
+- **Host:** NixOS container on [`shill`](../../shill.md)
+
+## Role
+
+- **matrix-synapse** — `server_name = "nul.ie"`, `public_baseurl = https://matrix.nul.ie`,
+  Element at `element.nul.ie` as the web client. Listens on `[::]:8008` (client + federation
+  resources, `x_forwarded`) with a localhost manhole on `:9000`. Registration and guest access
+  are disabled; uploads up to 1024M with dynamic thumbnails and URL previews enabled (previews
+  are limited to [middleman](middleman.md)'s addresses as the fetch proxy).
+- **heisenbridge** — IRC bridge, owner `@dev:nul.ie`, exclusive `@irc_*` user namespace.
+- **mautrix-whatsapp** — WhatsApp bridge (appservice `whatsapp2`, `!wa` commands).
+- **mautrix-meta** — two instances, `messenger` (`fbm2_*`, `!fbm`) and `instagram` (`ig_*`,
+  `!ig`), both with backfill enabled.
+- All three mautrix bridges use Postgres on [colony-psql](colony-psql.md) (URIs in their secret
+  env files), require end-to-end encryption by default, and double-puppet onto `nul.ie` via the
+  shared `doublepuppet.yaml` appservice registration (an age secret).
+- The firewall allows `8008` (Synapse) and `8009` besides netdata.
+
+## Network assignments
+
+<!-- assignments: chatterbox -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| chatterbox-ctr | internal | `10.100.2.5/24 gw 10.100.2.1` | `2a0e:97c0:4d2:12::5/64` | ams1.int.nul.ie |  |
+<!-- assignments-end -->
+
+## Notes
+
+- Synapse's real database config lives in the `chatterbox/synapse.yaml` age secret — options
+  only merge at the top level, so the base config carries a dummy `sqlite3` block to satisfy
+  the module defaults. The signing key is also an age secret.
+- `olm-3.2.16` is allowed via `permittedInsecurePackages` (a nixpkgs E2EE library issue).
+- The bridge services get `ffmpeg` on their `PATH` for GIF→video conversion.
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/shill/containers/chatterbox.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/chatterbox.nix) — container definition, Synapse settings and all bridge configuration
@@ -0,0 +1,43 @@
+# colony-psql
+
+The shared PostgreSQL instance for colony services. Rather than each service running its own
+database, the containers (and the `git` VM) connect here over the `ctrs` network.
+
+- **Source:** [`shill/containers/colony-psql.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/colony-psql.nix)
+- **Host:** NixOS container on [`shill`](../../shill.md)
+
+## Role
+
+- **PostgreSQL 14** with TCP/IP enabled, reachable from the whole colony (`10.100.0.0/16` and
+  `2a0e:97c0:4d2:10::/60`, `md5` auth). The firewall allows `5432`.
+- Local `peer` auth maps `postgres`, `root`, `netdata` and `dev` to the `postgres` superuser via
+  the ident map.
+- **netdata** with the Python PostgreSQL collector.
+- Consumers wait for the database to accept connections with the `lib.my.systemdAwaitPostgres`
+  helper (e.g. `sharry`, `atticd`, `mastodon-init-db`, and `middleman`'s nginx as a DNS
+  bootstrap hack).
+
+## Network assignments
+
+<!-- assignments: colony-psql -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| colony-psql-ctr (colony-psql) | internal | `10.100.2.4/24 gw 10.100.2.1` | `2a0e:97c0:4d2:12::4/64` | ams1.int.nul.ie |  |
+<!-- assignments-end -->
+
+The assignment also has the alt name `colony-psql` (no `-ctr` suffix), which is what consumers
+use as the database hostname.
+
+## Consumers
+
+- [object](object.md) — `sharry` and `hedgedoc` (and `atticd` when enabled) over
+  `colony-psql:5432`
+- [toot](toot.md) — Mastodon's database (Mastodon currently disabled)
+- [chatterbox](chatterbox.md) — the mautrix bridges (WhatsApp, Messenger, Instagram) via
+  Postgres URIs in their secret env files
+- `git` VM — Gitea
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/shill/containers/colony-psql.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/colony-psql.nix) — container definition and PostgreSQL configuration
@@ -0,0 +1,27 @@
+# gam
+
+A game-server container — the lightweight counterpart to the OCI game servers on `whale2`,
+running servers directly as NixOS services. Currently it runs a single Terraria server.
+
+- **Source:** [`shill/containers/gam.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/gam.nix)
+- **Host:** NixOS container on [`shill`](../../shill.md)
+
+## Role
+
+- **terraria** — dedicated server (`services.terraria`): world at
+  `/var/lib/terraria/NotWorld.wld`, auto-created large worlds, MOTD "sup gamers", UPnP off.
+  Extra settings (e.g. password) come from the `gam/terraria.conf` age secret used as the config
+  file. `openFirewall` is on, and `estuary` port-forwards TCP and UDP `7777` to this container.
+
+## Network assignments
+
+<!-- assignments: gam -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| gam-ctr | internal | `10.100.2.11/24 gw 10.100.2.1` | `2a0e:97c0:4d2:12::b/64` | ams1.int.nul.ie |  |
+<!-- assignments-end -->
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/shill/containers/gam.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/gam.nix) — container definition and the Terraria service
@@ -0,0 +1,62 @@
+# jackflix
+
+The media stack — acquisition, library, streaming and photos. Torrent traffic is routed through
+an AirVPN WireGuard tunnel so downloads only flow while the VPN is up.
+
+- **Source:** [`shill/containers/jackflix/`](../../../../../nixos/boxes/colony/vms/shill/containers/jackflix)
+  (`default.nix`, `networking.nix`)
+- **Host:** NixOS container on [`shill`](../../shill.md) (bind-mounts `/mnt/media` read-write)
+
+## Role
+
+| Service | Port | Purpose |
+| --- | --- | --- |
+| Jellyfin | `8096` | streaming, published as `jackflix.nul.ie` |
+| Transmission | `9091` | BitTorrent client (`transmission_4`), published as `torrents.nul.ie` (SSO) |
+| Jackett | `9117` | indexer aggregator, `jackett.nul.ie` (SSO) |
+| FlareSolverr | — | Cloudflare challenge solver for Jackett |
+| Radarr | `7878` | movies, `radarr.nul.ie` (SSO) |
+| Sonarr | `8989` | TV, `sonarr.nul.ie` (SSO) |
+| Jellyseerr (`seerr`) | `5055` | request portal, `gib.nul.ie` (`openFirewall` on) |
+| PhotoPrism | `2342` | photos, `photos.nul.ie`; password auth, sqlite DB, originals/import under `/mnt/media/photoprism` |
+| copyparty | `3923` | file sharing, `stuff.nul.ie`; serves `/mnt/media/public` (read-only to everyone) and `/priv` → `/mnt/media/stuff` (admin for `dev`), share creation, indexing (`e2dsa`/`e2t`), file-magic checks |
+
+All published through [middleman](middleman.md) as shown. A shared `media` group (gid 2000) plus
+`UMask=0002` on Radarr/Sonarr gives the apps coordinated access to the media volume.
+
+## Network assignments
+
+<!-- assignments: jackflix -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| jackflix-ctr | internal | `10.100.2.6/24 gw 10.100.2.1` | `2a0e:97c0:4d2:12::6/64` | ams1.int.nul.ie |  |
+<!-- assignments-end -->
+
+## VPN download path
+
+[`networking.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/jackflix/networking.nix)
+defines a `vpn` WireGuard netdev to **AirVPN NL** (key + PSK from age secrets, MTU 1320 per
+AirVPN, fwmark `42`, route table `51820`):
+
+- Policy routing keeps colony traffic on the main table (from/to `prefixes.all` v4/v6 rules at
+  priority 100, plus a default-route suppression) while everything else falls through to the VPN
+  table — so the services stay reachable on the `ctrs` network while outbound torrent traffic
+  exits via AirVPN. `DNSDefaultRoute` is disabled on `host0`; the VPN provides DNS.
+- `transmission` and `jackett` `bindsTo` `systemd-networkd-wait-online@vpn.service` — they only
+  run while the tunnel is up.
+- AirVPN forwards peer port `28457` to Transmission (`peer-port`); the firewall accepts it and
+  drops other new inbound TCP from `vpn`, while non-VPN input is limited to the service ports
+  (netdata, Transmission, Jackett, Radarr, Sonarr, Jellyfin, PhotoPrism) plus copyparty's `3923`
+  from the base config and Jellyseerr's `5055`.
+
+## Storage
+
+Media lives on the shared `/mnt/media` volume (bind-mounted read-write from `shill`); Transmission
+downloads into `/mnt/media/downloads/torrents` with a `.incomplete` dir, 28 MB/s up/down limits
+and a seed ratio limit of 2.0.
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/shill/containers/jackflix/default.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/jackflix/default.nix) — container definition and the media services
+- [`nixos/boxes/colony/vms/shill/containers/jackflix/networking.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/jackflix/networking.nix) — AirVPN WireGuard netdev, policy routing and VPN firewall rules
@@ -0,0 +1,94 @@
+# middleman
+
+The front-end reverse proxy for the colony's public web services — the single ingress that
+`estuary` DNATs HTTP/HTTPS (and Matrix federation on `:8448`) to. Terminates TLS with wildcard
+certificates it issues itself, provides nginx-sso for gated vhosts, and runs a librespeed
+backend.
+
+- **Source:** [`shill/containers/middleman/`](../../../../../nixos/boxes/colony/vms/shill/containers/middleman)
+  (`default.nix`, `vhosts.nix`)
+- **Host:** NixOS container on [`shill`](../../shill.md) (`my.containers` ephemeral nspawn on the
+  `ctrs` bridge; bind-mounts `/mnt/media` read-only for the static file vhosts)
+
+## Role
+
+- **nginx** reverse proxy with the `vts` and `fancyindex` modules, Brotli, kTLS, a proxy cache,
+  and a dynamic resolver pointed at `estuary` (`proxyResolveWhileRunning`) so upstreams re-resolve
+  at runtime. All vhosts are declared in
+  [`vhosts.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/middleman/vhosts.nix);
+  upstreams are addressed by their internal container/VM hostnames under `ams1.int.nul.ie`.
+  nginx is made to wait for `colony-psql` (`systemdAwaitPostgres`) before starting — a hack for
+  it getting stuck on DNS early in boot.
+- **ACME** — issues the wildcard certificates its own vhosts are served with (it is not a shared
+  CA for the other boxes):
+  - `ams1.int.nul.ie` + `*.ams1.int.nul.ie` via a lego `exec` DNS challenge that SSHes to
+    `pdns-file-records@estuary-vm` to add/remove records in the internal PowerDNS zone. This is
+    the default cert (`useACMEHost`) for vhosts on the internal domain.
+  - `nul.ie` + `*.nul.ie` + `*.s3.nul.ie` via the Cloudflare DNS provider. A `postRun` hook
+    `scp`s the renewed key/cert to the `mail` (Mailcow) VM and runs `mailcow-ssl-reload` there.
+  - Certs reload `nginx` on renewal; the `acme` group owns the secret files and `nginx` is a
+    member.
+- **nginx-sso** — single sign-on at `sso.nul.ie` (one instance, `generic`), with Google OAuth as
+  the default method and a simple username/password provider. Cookie domain is `.nul.ie`.
+  SSO-gated vhosts below include `server-generic.conf` / `location-generic.conf` from
+  `/etc/nginx/includes/sso/`.
+- **librespeed** — speed test, backend + frontend, published as `speed.nul.ie` and
+  `librespeed.ams1.int.nul.ie` (both proxy to `localhost:8989`).
+
+## Network assignments
+
+<!-- assignments: middleman -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| middleman-ctr | internal | `10.100.2.2/24 gw 10.100.2.1` | `2a0e:97c0:4d2:12::2/64` | ams1.int.nul.ie |  |
+<!-- assignments-end -->
+
+The firewall allows `http`, `https` and `8448` (Matrix federation). A small nftables SNAT rule
+rewrites outbound IPv6 to the container's own address on `host0`.
+
+## Published vhosts
+
+Everything is under `*.nul.ie` with the public wildcard cert unless noted; defaults applied to
+all vhosts are `onlySSL`, kTLS and HTTP/2. "SSO" = gated behind nginx-sso (`generic` instance).
+
+| Host | Upstream | Notes |
+| --- | --- | --- |
+| `nul.ie` (`_`, default) | static | landing page (`index.html`, CV PDF, SSH pubkey); serves Matrix `.well-known`s and redirects `webfinger`/`nodeinfo`/`host-meta` → `toot.nul.ie`, `atproto-did` → `pds.nul.ie`; `forceSSL` (plain HTTP redirects to HTTPS) |
+| `localhost` | — | loopback-only VTS status page at `/status` (scraped by netdata); plain HTTP |
+| `sso.nul.ie` | `localhost:8082` | nginx-sso endpoint |
+| `netdata-colony.nul.ie` | `<host>.ams1.int.nul.ie:19999` | netdata fan-out over `vm`, `fw`, `ctr`, `oci`, `http`, `jackflix-ctr`, `chatterbox-ctr`, `colony-psql-ctr`; **SSO** |
+| `pass.nul.ie` | `vaultwarden-ctr:8080` | [vaultwarden](vaultwarden.md); `/notifications/hub` proxied with websockets |
+| `matrix.nul.ie` | `chatterbox-ctr:8008` | [chatterbox](chatterbox.md) Synapse client + federation; also listens on `:8448` as federation `default_server`; `= /` redirects to Element; serves Matrix `.well-known`s |
+| `element.nul.ie` | static `element-web` | Element configured for the `nul.ie` homeserver |
+| `torrents.nul.ie` | `jackflix-ctr:9091` | Transmission ([jackflix](jackflix.md)); **SSO** |
+| `jackett.nul.ie` | `jackflix-ctr:9117` | **SSO** |
+| `radarr.nul.ie` | `jackflix-ctr:7878` | **SSO**; websockets |
+| `sonarr.nul.ie` | `jackflix-ctr:8989` | **SSO**; websockets |
+| `gib.nul.ie` | `jackflix-ctr:5055` | Jellyseerr requests |
+| `jackflix.nul.ie` | `jackflix-ctr:8096` | Jellyfin; `/socket` websockets; `/` redirects to `/web/` |
+| `toot.nul.ie` | `toot-ctr:80` | Mastodon — **upstream currently disabled**, see [toot](toot.md) |
+| `pds.nul.ie` | `toot-ctr:3000` | Bluesky PDS ([toot](toot.md)); websockets |
+| `share.nul.ie` | `object-ctr:9090` | Sharry ([object](object.md)); websockets |
+| `stuff.nul.ie` | `jackflix-ctr:3923` | copyparty |
+| `public.nul.ie` (+ alias `p.nul.ie`) | static `/mnt/media/public` | fancyindex file listing; `addSSL` so plain HTTP also works |
+| `mc-map.nul.ie` | `simpcraft-oci:8100` | Minecraft map (OCI container on `whale2`) |
+| `mc-rail.nul.ie` | `simpcraft-oci:3876` | Minecraft railway map |
+| `mc-map-kink.nul.ie` | `kinkcraft-oci:8100` | Minecraft map |
+| `speed.nul.ie` | `localhost:8989` | librespeed |
+| `librespeed.ams1.int.nul.ie` | `localhost:8989` | librespeed on the internal domain (internal wildcard cert) |
+| `md.nul.ie` | `object-ctr:3000` | HedgeDoc; websockets |
+| `pb.nul.ie` | `object-ctr:8088` | wastebin |
+| `photos.nul.ie` | `jackflix-ctr:2342` | PhotoPrism; websockets |
+| `pront.nul.ie` | `stream-hi.h.nul.ie:5000` | OctoPrint on the home network; `/webcam/` → `:5050`; **SSO** |
+| `hass.nul.ie` | `hass-ctr.h.nul.ie:8123` | Home Assistant (home network); websockets |
+| `hass-john.nul.ie` | `john-valorant-tun.ams1.int.nul.ie:8123` | remote Home Assistant over the point-to-point tunnel; websockets |
+| `minio.nul.ie` | `object-ctr:9001` | MinIO console; `/ws` websockets |
+| `s3.nul.ie` (+ `*.s3.nul.ie`) | `object-ctr:9000` | MinIO S3 API (virtual-host style via the `*.s3` wildcard cert); `/gitea/packages/` has a hack forcing the correct `Content-Type` for Docker image manifests |
+| `nix-cache.nul.ie` | `object-ctr:5000` | Harmonia Nix binary cache; `.narinfo`/`nar/`/`serve/` paths get immutable `Cache-Control`/`Expires` headers |
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/shill/containers/middleman/default.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/middleman/default.nix) — container definition: nginx, ACME, nginx-sso, librespeed, secrets
+- [`nixos/boxes/colony/vms/shill/containers/middleman/vhosts.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/middleman/vhosts.nix) — all virtual hosts, the SSO include helpers, and the `.well-known` tree
+- [`nixos/boxes/colony/vms/shill/containers/middleman/default.html`](../../../../../nixos/boxes/colony/vms/shill/containers/middleman/default.html) — default vhost landing page
@@ -0,0 +1,43 @@
+# object
+
+Object storage and the Nix binary cache, plus a few small self-hosted web apps (Sharry,
+HedgeDoc, wastebin).
+
+- **Source:** [`shill/containers/object.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/object.nix)
+- **Host:** NixOS container on [`shill`](../../shill.md) (bind-mounts `/mnt/minio` and
+  `/mnt/nix-cache` read-write)
+
+## Role
+
+| Service | Port | Purpose |
+| --- | --- | --- |
+| MinIO | `9000` (S3) / `9001` (console) | S3-compatible object storage, `s3.nul.ie` + `*.s3.nul.ie` (virtual-host style via `MINIO_DOMAIN`), console at `minio.nul.ie`; region `eu-central-1`; data on the `/mnt/minio` XFS volume |
+| Harmonia | `5000` | Nix binary cache at `nix-cache.nul.ie` — `harmonia-dev` cache serves `shill`'s `/nix/store` out of a dedicated store view rooted at `/var/lib/harmonia` (bind-mounted from `/mnt/nix-cache`), signed with the `nix-cache.key` secret; a `harmonia` user with authorized keys exists for cache pushes; priority 30 |
+| Sharry | `9090` | file sharing at `share.nul.ie`; Postgres on [colony-psql](colony-psql.md), files stored in the `share` MinIO bucket; fixed `dev` account + invite signup; mail via `mail.nul.ie`; max share size 128G |
+| HedgeDoc | `3000` | collaborative markdown notes at `md.nul.ie`; Postgres on [colony-psql](colony-psql.md); anonymous edits but no anonymous notes, email login, no open email registration |
+| wastebin | `8088` | pastebin at `pb.nul.ie` |
+| atticd | `8069` | **currently disabled** (`services.atticd.enable = false`) — an alternative Nix cache that would store locally and sit behind `nix-cache.nul.ie`; config (including the `object/atticd.env` secret) is kept around |
+
+Everything public is fronted by [middleman](middleman.md) (see its vhost table). The
+`minio-client` is installed and the user's `~/.mc/config.json` points at an age-secret config.
+`minio-2025-10-15T17-29-55Z` is allowlisted via `permittedInsecurePackages` (flagged as a TODO).
+
+## Network assignments
+
+<!-- assignments: object -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| object-ctr | internal | `10.100.2.7/24 gw 10.100.2.1` | `2a0e:97c0:4d2:12::7/64` | ams1.int.nul.ie |  |
+<!-- assignments-end -->
+
+## Backing services
+
+- [colony-psql](colony-psql.md) — Sharry and HedgeDoc databases (atticd too, when enabled).
+- MinIO buckets back other boxes' services: Gitea LFS/packages (with the `middleman` MIME hack
+  for Docker manifests), Mastodon's `mastodon` bucket and the Bluesky PDS `pds` bucket on
+  [toot](toot.md), and Sharry's `share` bucket.
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/shill/containers/object.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/object.nix) — container definition and all services
@@ -0,0 +1,39 @@
+# qclk
+
+A WireGuard management appliance for the `qclk` network — it terminates the `management`
+tunnel and routes/NATs the `qclk` prefix. No service daemon is currently defined in the config;
+the container provides the network plumbing and opens the API port.
+
+- **Source:** [`shill/containers/qclk/`](../../../../../nixos/boxes/colony/vms/shill/containers/qclk)
+  (`default.nix`)
+- **Host:** NixOS container on [`shill`](../../shill.md)
+
+## Role
+
+- **WireGuard `management` interface** — listens on UDP `51821` (`lib.my.c.colony.qclk.wgPort`,
+  allowed through the firewall; `estuary` port-forwards it here) with the private key from the
+  `qclk/wg.key` age secret. Managed devices are static peers, each pinned to its own address in
+  the `qclk` prefix (`10.100.4.0/24`); the peer list currently has a single entry (host 2).
+- **Routing/NAT** — the container itself is host 1 of the `qclk` prefix. `shill` routes
+  `10.100.4.0/24` to this container, and outbound traffic from `host0` into `management` is
+  SNATed to the container's `qclk` address. Forwarding into `management` is accepted from the
+  AS211024 trusted IPv4 ranges (`lib.my.c.as211024.trusted.v4`).
+- **API port** — TCP `8080` is accepted on the `management` interface (`apiPort`), but note
+  `services = { }`: whatever serves the qclk API is not defined in this configuration today.
+
+## Network assignments
+
+<!-- assignments: qclk -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| qclk-ctr | internal | `10.100.2.10/24 gw 10.100.2.1` | `2a0e:97c0:4d2:12::a/64` | ams1.int.nul.ie |  |
+| qclk | qclk | `10.100.4.1/24` | — | — |  |
+<!-- assignments-end -->
+
+Two assignments: `internal` on the `ctrs` network like the other containers, and `qclk` — host 1
+of the `qclk` prefix on the `management` WireGuard interface (IPv4 only, no DNS name).
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/shill/containers/qclk/default.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/qclk/default.nix) — container definition: WireGuard netdev, peer list and firewall rules
@@ -0,0 +1,48 @@
+# toot
+
+Federated-social container. Despite the name, the only service actually running is a **Bluesky
+PDS** — the Mastodon instance ("toots") is **currently disabled**.
+
+- **Source:** [`shill/containers/toot.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/toot.nix)
+- **Host:** NixOS container on [`shill`](../../shill.md)
+
+## Role
+
+- **bluesky-pds** — the active service. `pds.nul.ie` on port `3000`, fronted by
+  [middleman](middleman.md) (which also redirects `/.well-known/atproto-did` here). Invites
+  required; blob store is the `pds` bucket on [object](object.md)'s MinIO (`s3.nul.ie`,
+  `eu-central-1`), upload limit 50 MiB; federation settings point at the stock Bluesky
+  infrastructure (`plc.directory`, `api.bsky.app`, `mod.bsky.app`, `bsky.network` crawlers).
+  Secrets (including the S3 credentials) come from the `toot/pds.env` age secret. Email from
+  `pds@nul.ie`.
+- **mastodon** — **disabled** (`services.mastodon.enable = false`). The full config is still
+  present: `LOCAL_DOMAIN = nul.ie` with `WEB_DOMAIN = toot.nul.ie`, Postgres on
+  [colony-psql](colony-psql.md), local Redis, SMTP via `mail.nul.ie`, media in the `mastodon`
+  MinIO bucket (`S3_ALIAS_HOST = mastodon.s3.nul.ie`), 4 streaming processes, and media
+  auto-cleanup after 30 days. [middleman](middleman.md) still proxies `toot.nul.ie` →
+  `toot-ctr:80` and redirects the `webfinger`/`nodeinfo`/`host-meta` well-knowns there, but with
+  the service off those endpoints are dead. The config notes the removed `otpSecretFile` option
+  would need addressing before Mastodon can come back.
+
+## Network assignments
+
+<!-- assignments: toot -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| toot-ctr | internal | `10.100.2.8/24 gw 10.100.2.1` | `2a0e:97c0:4d2:12::8/64` | ams1.int.nul.ie |  |
+<!-- assignments-end -->
+
+The firewall allows `http` (the Mastodon nginx vhost) and the PDS port `3000` besides netdata.
+
+## Notes
+
+- The local nginx still carries the Mastodon virtual host (`toot.nul.ie`) with proxy-header
+  overrides for being behind `middleman` — part of the preserved-but-disabled Mastodon setup.
+- `mastodon-init-dirs` appends the S3 secret key to Mastodon's `.secrets_env` (the module has no
+  option for a secret-key file), and `mastodon-init-db` waits for `colony-psql` — moot while
+  Mastodon is disabled.
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/shill/containers/toot.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/toot.nix) — container definition; active PDS config and the preserved (disabled) Mastodon config
@@ -0,0 +1,37 @@
+# vaultwarden
+
+[Vaultwarden](https://github.com/dani-garcia/vaultwarden), a Bitwarden-compatible password
+manager, published as `pass.nul.ie` through [middleman](middleman.md).
+
+- **Source:** [`shill/containers/vaultwarden.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/vaultwarden.nix)
+- **Host:** NixOS container on [`shill`](../../shill.md)
+
+## Role
+
+- **vaultwarden** — HTTP on `[::]:8080`, WebSocket notifications on `3012` (both allowed through
+  the firewall). Web vault enabled, signups disabled, Bitwarden push notifications enabled
+  (`PUSH_ENABLED`). `DOMAIN` is `https://pass.nul.ie`.
+- **SMTP** via `mail.nul.ie:587` (STARTTLS) as `pass@nul.ie`; credentials and other sensitive
+  settings come from the `vaultwarden/config.env` age secret.
+- **Backups** — a `borgbackup` job pushes `/var/lib/vaultwarden` to rsync.net
+  (`zh2855@zh2855.rsync.net:borg/vaultwarden2`), repokey-encrypted (passphrase and SSH key from
+  secrets), `zstd,10` compression, keeping 7 daily / 4 weekly / all monthly archives.
+
+## Network assignments
+
+<!-- assignments: vaultwarden -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| vaultwarden-ctr | internal | `10.100.2.3/24 gw 10.100.2.1` | `2a0e:97c0:4d2:12::3/64` | ams1.int.nul.ie |  |
+<!-- assignments-end -->
+
+## Persistence
+
+`/var/lib/vaultwarden` is persisted through `my.tmproot.persistence` — like the other
+`shill` containers the root is ephemeral and real state lives under `/persist` (bind-mounted
+from the host's `/persist/containers/vaultwarden`).
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/shill/containers/vaultwarden.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/vaultwarden.nix) — container definition, service config and the borgbackup job
@@ -0,0 +1,40 @@
+# waffletail
+
+The colony Tailscale node: a subnet router and exit node that advertises the colony prefixes
+into the tailnet.
+
+- **Source:** [`shill/containers/waffletail.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/waffletail.nix)
+- **Host:** NixOS container on [`shill`](../../shill.md)
+
+## Role
+
+- **tailscale** — joins via an auth key from secrets (`tailscale-auth.key`) against the
+  self-hosted Headscale control plane at `hs.nul.ie`. Runs with `--netfilter-mode=off` (firewall
+  is managed by the repo's own nftables rules), `--advertise-exit-node`, and
+  `--advertise-routes` covering the whole colony — `10.100.0.0/16` and `2a0e:97c0:4d2:10::/60`.
+  Does **not** accept routes itself. Listens on UDP `41641` (`openFirewall`), which `estuary`
+  port-forwards to this container.
+- `shill` routes the Tailscale prefixes (`100.64.0.0/10`, `fd7a:115c:a1e0::/48`) to this
+  container, so colony hosts can reach tailnet clients and vice versa.
+- nftables: `tailscale0` is a trusted interface; forwarding from `host0` into Tailscale is
+  allowed for the colony source ranges, and tailnet-sourced traffic leaving via `host0` is
+  SNATed to the container's colony addresses (except when destined to the colony ranges
+  themselves).
+
+## Network assignments
+
+<!-- assignments: waffletail -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| waffletail-ctr | internal | `10.100.2.9/24 gw 10.100.2.1` | `2a0e:97c0:4d2:12::9/64` | ams1.int.nul.ie |  |
+| waffletail | tailscale | `100.64.0.5/32` | `fd7a:115c:a1e0::5/128` | — |  |
+<!-- assignments-end -->
+
+Two assignments: `internal` on the `ctrs` network like the other containers, and `tailscale` —
+its addresses on the tailnet itself (host 5 of `100.64.0.0/10` and `fd7a:115c:a1e0::/48`; no
+DNS name).
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/shill/containers/waffletail.nix`](../../../../../nixos/boxes/colony/vms/shill/containers/waffletail.nix) — container definition, Tailscale setup and forward/NAT rules
@@ -0,0 +1,84 @@
+# whale2
+
+The colony podman/OCI host, dedicated to game servers (kept off `shill` so
+container churn and resource use stay isolated).
+
+- **Source:** [`nixos/boxes/colony/vms/whale2/`](../../../nixos/boxes/colony/vms/whale2)
+  (`default.nix`, `valheim.nix`, `minecraft/`, `enshrouded.nix`)
+- **Host:** VM on `colony`
+- **nixpkgs:** `mine`
+
+## Role
+
+- Runs OCI containers via podman (`virtualisation.oci-containers`, netavark
+  backend, `firewall_driver = "none"` so podman doesn't fight `my.firewall`).
+- Each game server gets its own routable address on the `colony` netavark
+  network (defined in `/etc/containers/networks/colony.json`), which is backed
+  by the `oci` interface and the `prefixes.oci` v4/v6 ranges; per-game
+  addresses come from `extraAssignments` (`valheim-oci`, `simpcraft-oci`, …)
+  and are passed to podman with `--network=colony:ip=…` (`lib.my.dockerNetAssignment`).
+- `estuary` forwards the game ports in (see `firewallForwards`), so the
+  servers are reachable on the public IP as well as directly over IPv6.
+- `/var/lib/containers` is a dedicated XFS disk (project quotas).
+
+## Network assignments
+
+<!-- assignments: whale2 -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| whale-vm (oci) | internal | `94.142.241.226/32` | `2a0e:97c0:4d2:11::3/64` | ams1.int.nul.ie |  |
+| whale-vm-oci | oci | `10.100.3.1/24` | `2a0e:97c0:4d2:13::1/64` | ams1.int.nul.ie |  |
+| whale-vm-routing | routing | `10.100.1.3/24 gw 10.100.1.1` | — | ams1.int.nul.ie |  |
+<!-- assignments-end -->
+
+## Game servers
+
+The OCI containers are documented here (they have no pages of their own).
+Addresses are the per-container `extraAssignments` on the `oci` network;
+ports are the public ones forwarded by `estuary`.
+
+| Container | Address (v4 / v6 host) | Ports | Status |
+|---|---|---|---|
+| `valheim` | `10.100.3.2` / `2a0e:97c0:4d2:13::2` | `2456-2457`/udp | running |
+| `simpcraft` | `10.100.3.3` / `2a0e:97c0:4d2:13::3` | `25565` tcp+udp | running |
+| `simpcraft-staging` | `10.100.3.4` / `2a0e:97c0:4d2:13::4` | `25566` tcp | **disabled** (commented out) |
+| `enshrouded` | `10.100.3.5` / `2a0e:97c0:4d2:13::5` | `15636-15637`/udp | **disabled** (`enshrouded.nix` not imported) |
+| `kevcraft` | `10.100.3.6` / `2a0e:97c0:4d2:13::6` | `25567` tcp+udp | running |
+| `kinkcraft` | `10.100.3.7` / `2a0e:97c0:4d2:13::7` | `25568` tcp+udp | running |
+| `graeme` | `10.100.3.8` / `2a0e:97c0:4d2:13::8` | `25569` tcp+udp | running |
+
+- **valheim** ([`valheim.nix`](../../../nixos/boxes/colony/vms/whale2/valheim.nix)) —
+  `lloesche/valheim-server`, public server "amogus sus", world `simpland2`,
+  allow-listed Steam IDs, password from agenix.
+- **simpcraft** ([`minecraft/`](../../../nixos/boxes/colony/vms/whale2/minecraft)) —
+  `itzg/minecraft-server` (self-built `git.nul.ie/dev/craftblock` image),
+  Modrinth "Simpcraft" modpack, whitelist + ops, 8 GiB heap.
+  **simpcraft-staging** is the same setup pinned to an older pack version,
+  currently commented out.
+- **kevcraft** — vanilla Minecraft 1.20.1, 4 GiB heap, extra op.
+- **kinkcraft** — same Simpcraft modpack as `simpcraft`, 6 GiB heap.
+- **graeme** — vanilla Minecraft on hard difficulty with its own whitelist.
+- **enshrouded** ([`enshrouded.nix`](../../../nixos/boxes/colony/vms/whale2/enshrouded.nix)) —
+  `sknnr/enshrouded-dedicated-server` ("UWUshrouded"); the file exists but is
+  commented out of `whale2`'s `imports`, so the server is down (its forwards
+  and DNS records remain).
+
+The Minecraft containers share one whitelist/ops list and agenix env file
+(`whale2/simpcraft.env`, which also carries the RCON password).
+
+## Backups
+
+A local borg job (`services.borgbackup.jobs.simpcraft`) archives the
+`simpcraft` world every ~15 minutes (offset from the usual 5-minute autosave
+ticks) into `/var/lib/containers/backup/simpcraft`, using `mcrcon` to
+`save-off`/`save-on` around each run. Retention is short (12 h + 48 hourly) —
+this is for quick world rollback, not disaster recovery (the `oci` LV itself
+is covered by `colony`'s `borgthin`).
+
+## Notable config files
+
+- [`nixos/boxes/colony/vms/whale2/default.nix`](../../../nixos/boxes/colony/vms/whale2/default.nix) — VM config, podman/netavark setup, `extraAssignments`.
+- [`nixos/boxes/colony/vms/whale2/valheim.nix`](../../../nixos/boxes/colony/vms/whale2/valheim.nix) — Valheim server.
+- [`nixos/boxes/colony/vms/whale2/minecraft/default.nix`](../../../nixos/boxes/colony/vms/whale2/minecraft/default.nix) — the Minecraft servers + world backup job.
+- [`nixos/boxes/colony/vms/whale2/enshrouded.nix`](../../../nixos/boxes/colony/vms/whale2/enshrouded.nix) — Enshrouded server (disabled, not imported).
@@ -0,0 +1,81 @@
+# Home site
+
+The home network (domain `h.nul.ie`): a redundant pair of routers in front of a VM host, an
+NVMe-oF storage target, an IoT container host, and a workstation. The two routers — `river` (a VM)
+and `stream` (a physical box) — are built from one shared
+[`routing-common`](../../../nixos/boxes/home/routing-common) definition as an active/backup VRRP
+pair, and everything clients touch (gateway, DNS) is a floating VIP that follows the master.
+
+- **Source:** [`nixos/boxes/home/`](../../../nixos/boxes/home)
+
+## Hierarchy
+
+```
+h.nul.ie
+├── palace (physical VM host — AMD, 100G, SR-IOV)
+│   ├── river ── primary router VM (PPPoE / Digiweb WAN)
+│   ├── cellar ─ NVMe-oF / SPDK storage target VM
+│   └── sfh ──── container host VM ("services for home")
+│       ├── hass ── Home Assistant + Frigate + MQTT (container)
+│       └── unifi ─ UniFi controller (container)
+├── stream (physical secondary router — Virgin Media WAN)
+└── castle (workstation / gaming desktop — netboot, NVMe-oF root)
+```
+
+## Machines
+
+| Box | Role | Host | Page |
+|---|---|---|---|
+| `palace` | VM host | physical | [palace.md](palace.md) |
+| `river` | Primary router (VRRP pair with `stream`) | VM on `palace` | [river.md](river.md) |
+| `stream` | Secondary router (VRRP pair with `river`) | physical | [stream.md](stream.md) |
+| `cellar` | NVMe-oF / SPDK storage target | VM on `palace` | [cellar.md](cellar.md) |
+| `sfh` | NixOS container host | VM on `palace` | [sfh.md](sfh.md) |
+| `castle` | Workstation / gaming desktop | physical | [castle.md](castle.md) |
+| `hass` | Home Assistant + Frigate + MQTT | container on `sfh` | [sfh/containers/hass.md](sfh/containers/hass.md) |
+| `unifi` | UniFi controller | container on `sfh` | [sfh/containers/unifi.md](sfh/containers/unifi.md) |
+
+## Router VIPs
+
+Clients never use a router's real address: `kea` (DHCP) and `radvd` (RAs) hand out the per-VLAN
+floating VIPs as both gateway and DNS server, and `keepalived` moves them between `river` and
+`stream`. They are declared as pseudo-systems (`router-hi` / `router-lo` / `router-ut`) in
+`routing-common`'s `extraAssignments`, with addresses from `lib.my.c.home.vips`
+([`lib/constants.nix`](../../../lib/constants.nix)); `pdns-recursor` binds them on both routers
+(`ip_nonlocal_bind`) so DNS follows the master instead of relying on client resolver timeouts.
+
+| Pseudo-system | Network | IPv4 | IPv6 |
+|---|---|---|---|
+| `router-hi` | `hi` (VLAN 100) | `192.168.71.254/22` | `2a0e:97c0:4d0:1::ffff/64` |
+| `router-lo` | `lo` (VLAN 110) | `192.168.79.254/21` | `2a0e:97c0:4d0:2::ffff/64` |
+| `router-ut` | `untrusted` (VLAN 120) | `192.168.80.254/24` | `2a0e:97c0:4d0:3::ffff/64` |
+
+Notes:
+
+- The IPv6 gateway clients learn from RAs is the link-local `fe80::1` on each VLAN (a
+  `virtual_ipaddress_excluded` VIP; `radvd` advertises from it), not the global VIP above.
+- `keepalived` also floats a VIP on the `as211024` mesh interface (`10.100.50.4`,
+  `2a0e:97c0:4df:0:1::ffff`) — it has no pseudo-system because nothing client-facing uses it.
+- The `untrusted` VLAN is an exception to "DNS follows the master": its DHCP/RA options hand out
+  Cloudflare resolvers (`1.1.1.1` / `2606:4700:4700::1111`), not the VIP.
+
+## Networks
+
+| Network | VLAN | IPv4 | IPv6 | MTU |
+|---|---|---|---|---|
+| `core` | native | `192.168.64.0/24` | — | 1500 |
+| `hi` | 100 | `192.168.68.0/22` | `2a0e:97c0:4d0:1::/64` | 9000 |
+| `lo` | 110 | `192.168.72.0/21` | `2a0e:97c0:4d0:2::/64` | 1500 |
+| `untrusted` | 120 | `192.168.80.0/24` | `2a0e:97c0:4d0:3::/64` | 1500 |
+| `wan` (stream) | 130 | DHCP public lease + modem mgmt `192.168.0.0/24` | — | 1500 |
+| `pon-isp` (river) | 10 | PPPoE transport (no L3) | — | 1508 |
+| `wan-pon-ont` (river) | 140 | ONT mgmt `192.168.100.0/24` | — | 1500 |
+
+The logical network map lives in [networking.md](../../networking.md).
+
+## Switch fabric
+
+The boxes hang off three hand-configured switches — `jim` and `dave` (MikroTik, RouterOS) and
+`brian` (Ubiquiti, UniFi) — which are **not** managed by this flake. The physical topology, VLAN
+map, the Digiweb WAN path (trunked VLAN 10 + PVID 140 at the ONT edge), and the multi-ONT plan are
+documented in [switches.md](switches.md).
@@ -0,0 +1,49 @@
+# castle
+
+The home workstation / gaming desktop. Diskless-style: it netboots from `river` and keeps its
+root storage on NVMe-oF volumes from `cellar`.
+
+- **Source:** [`nixos/boxes/home/castle/`](../../../nixos/boxes/home/castle) (`default.nix`)
+- **Host:** physical
+
+## Role
+
+- AMD desktop running the GUI stack (`my.gui.enable`, Sway/Wayland via home-manager), PipeWire
+  (low-latency `quantum 128`, EasyEffects, jacktrip), Bluetooth, Thunderbolt (`bolt`).
+- **Netboot client** (`my.netboot.client.enable`): the firmware iPXE-boots off the 2.5G NIC —
+  kea's client-class for `castle` matches `et2.5g`'s MAC (`c8:7f:54:6e:17:0f`) and points at
+  `boot.h.nul.ie` on [`river`](river.md).
+- **Root on NVMe-oF**: `/nix`, `/persist` and `/home` are `/dev/nvmeof/*` LVs on the
+  `nqn.2016-06.io.spdk:castle` namespace from [`cellar`](cellar.md) (`my.nvme.boot`,
+  `192.168.68.80`, RDMA). The initrd brings up `et100g`/`lan-hi` plus `roceBootModules` to reach
+  it, and the running system keeps `KeepConfiguration=static` on `lan-hi` so networkd never drops
+  the NVMe-oF address. The root itself is a 24 GiB tmpfs (`my.tmproot`).
+- Local virtualisation: `libvirtd` + `virt-manager` are enabled and the IOMMU is on
+  (`amd_iommu=on`), but no VFIO/GPU-passthrough is configured in the box config today.
+- Both firewalls are off (`networking.firewall.enable` and `my.firewall.enable` — it's a trusted
+  desktop on `hi`).
+- Misc: `binfmt` emulation for `aarch64-linux`/`armv7l-linux`, `recursive-nix`, Wireshark,
+  `rdma-core`/`qperf` for the RoCE link. A `drm-amd-display` flicker patch sits commented out in
+  `kernelPatches`.
+
+## Network assignments
+
+<!-- assignments: castle -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| castle | hi | `192.168.68.40/22 gw 192.168.71.254` | `2a0e:97c0:4d0:1::3:1/64` | h.nul.ie |  |
+<!-- assignments-end -->
+
+## Networking
+
+- `et100g` (100G, MTU 9000) carries `lan-hi` (the `hi` assignment, statically `.40` — also pinned
+  by a kea reservation on its MAC) and `lan-lo`.
+- `lan-lo` is a secondary leg: DHCPv4 with `UseGateway`/`UseDNS` off and RAs accepted with
+  gateway/DNS use off — present for reaching `lo` devices, never a default route.
+- `et2.5g` (netboot) and `et10g` are renamed but carry no network config.
+
+## Notable config files
+
+- [`nixos/boxes/home/castle/default.nix`](../../../nixos/boxes/home/castle/default.nix) — box
+  config: netboot/NVMe-oF boot, 100G networking, GUI/audio, virtualisation.
@@ -0,0 +1,53 @@
+# cellar
+
+The home storage target. A VM on `palace` that drives three passed-through NVMe disks with SPDK
+and exports them over NVMe-oF/RDMA — `river`, `sfh` and `castle` all run their root storage off
+it.
+
+- **Source:** [`nixos/boxes/home/palace/vms/cellar/`](../../../nixos/boxes/home/palace/vms/cellar)
+  (`default.nix`, `spdk.nix`)
+- **Host:** VM on `palace`
+
+## Role
+
+- Runs an **SPDK userspace target** (`my.spdk`, [`spdk.nix`](../../../nixos/boxes/home/palace/vms/cellar/spdk.nix)):
+  the kernel `nvme` driver is blacklisted so SPDK can claim the three NVMe controllers directly
+  (host BDFs `41:00.0`–`43:00.0`, attached in the guest as `02:00.0`–`04:00.0`).
+- Builds a **RAID-0** (`NVMeRaid`, 64 KiB strip) across the three drives and exports one
+  partition per consumer as an **NVMe-oF subsystem over RDMA** (port 4420) on the `hi` network:
+
+  | Bdev | NQN | Consumer |
+  |---|---|---|
+  | `NVMeRaidp1` | `nqn.2016-06.io.spdk:river` | [`river`](river.md) |
+  | `NVMeRaidp2` | `nqn.2016-06.io.spdk:castle` | [`castle`](castle.md) |
+  | `NVMeRaidp3` | `nqn.2016-06.io.spdk:sfh` | [`sfh`](sfh.md) |
+
+  Each subsystem is pinned to its consumer's `hostnqn` (the `my.nvme.uuid` on the client side).
+- `spdk-tgt` is ordered after `lan-hi` is online; the RDMA listener binds the `hi` address
+  (`192.168.68.80:4420`). The VM itself is pinned to NUMA node 1 on `palace` and gets SR-IOV VF 0.
+- `netdata` (port 19999 allowed in the firewall) and `fstrim`.
+- Deploy address `192.168.68.80` (`my.deploy.node.hostname`).
+
+## Network assignments
+
+<!-- assignments: cellar -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| cellar | hi | `192.168.68.80/22 gw 192.168.71.254` | `2a0e:97c0:4d0:1::4:1/64` | h.nul.ie |  |
+<!-- assignments-end -->
+
+## Notes
+
+- The `ublk_*` calls in `my.spdk.debugCommands` are only a debugging aid — they create a local
+  ublk device so the RAID can be mounted and inspected on `cellar` itself. Client exports are the
+  `nvmf` subsystems above.
+
+## Notable config files
+
+- [`nixos/boxes/home/palace/vms/cellar/default.nix`](../../../nixos/boxes/home/palace/vms/cellar/default.nix) —
+  box config (assignment, networking, netdata).
+- [`nixos/boxes/home/palace/vms/cellar/spdk.nix`](../../../nixos/boxes/home/palace/vms/cellar/spdk.nix) —
+  SPDK target: RAID-0, NVMe-oF/RDMA subsystems.
+- [`nixos/boxes/home/palace/vms/default.nix`](../../../nixos/boxes/home/palace/vms/default.nix) —
+  the VM definition on `palace` (VF 0, NVMe passthrough, NUMA pinning).
@@ -0,0 +1,59 @@
+# palace
+
+The physical VM host for the home site. Runs the `river`, `cellar` and `sfh` VMs and feeds them
+SR-IOV VFs, PCI NVMe drives and LVM disks.
+
+- **Source:** [`nixos/boxes/home/palace/default.nix`](../../../nixos/boxes/home/palace/default.nix)
+  (VM definitions in [`palace/vms/default.nix`](../../../nixos/boxes/home/palace/vms/default.nix))
+- **Host:** physical
+
+## Role
+
+- Home hypervisor: VMs are declared in `my.vms.instances`
+  ([`palace/vms/default.nix`](../../../nixos/boxes/home/palace/vms/default.nix)); disks are LVs in
+  the `main` thin pool (`services.lvm.boot.thin.enable`).
+- AMD box (`kvm-amd`, `amd_iommu=on`, microcode updates); the kernel is built with
+  `ACPI_APEI_PCIEAER`/`PCIEAER` for the PCIe passthrough work below.
+- Deploy address `192.168.68.22` (`my.deploy.node.hostname`).
+
+## Network assignments
+
+<!-- assignments: palace -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| palace-core | core | `192.168.64.20/24` | — | h.nul.ie |  |
+| palace | hi | `192.168.68.22/22 gw 192.168.71.254` | `2a0e:97c0:4d0:1::2:1/64` | h.nul.ie |  |
+<!-- assignments-end -->
+
+## Networking
+
+- 100G `et100g` (mlx5, MTU 9000) uplinks to the `dave` switch and carries `lan-hi` (VLAN 100,
+  the `hi` assignment). A udev rule creates **4 SR-IOV VFs** on the PF, distributed in
+  `50-et100g`'s `[SR-IOV]` sections: VF 0 → `cellar` (VLAN hi), VF 1 → `river` (no VLAN — river
+  tags all of its own VLANs, including both WAN VLANs, on this trunk), VF 2 → `sfh` (VLAN hi),
+  VF 3 → `sfh`'s container MACVLAN parent (VLAN hi).
+- `lan-core` is a bridge with the `core` assignment (`192.168.64.20`, no gateway); the 1G
+  `lan-core-phy` and the `lan-lo-phy` VLAN ride on it. `lan-lo` is a second, L3-less bridge used
+  for VM netboot and `lo` clients.
+- The 1G `et1g0` (igb) is exported to `river` as a passthru-mode macvtap (`vm-et1g0`) — river sees
+  it as `wan-old`.
+
+## VMs
+
+| VM | vCPUs | RAM | Passthrough | Notes |
+|---|---|---|---|---|
+| `cellar` | 8c × 2t | 16 GiB | VF 0 (`44:00.1`); NVMe `41:00.0`–`43:00.0` | pinned to NUMA node 1; split IRQ chip + vIOMMU |
+| `river` | 3c × 2t | 4 GiB | VF 1 (`44:00.2`); macvtap `vm-et1g0` | only an ESP local disk (plus an installer ISO) — root is NVMe-oF from `cellar` |
+| `sfh` | 8c × 2t | 32 GiB | VF 2 (`44:00.3`), VF 3 (`44:00.4`); two USB host ports | no boot disk — netboots; gets the `hdds/frigate` LV |
+
+Boot ordering is enforced with systemd dependencies: `vm@river` waits for `cellar`'s SSH port (and
+for the `vm-et1g0` device), and `vm@sfh` waits for `river` — storage first, then the router, then
+everything that boots off both.
+
+## Notable config files
+
+- [`nixos/boxes/home/palace/default.nix`](../../../nixos/boxes/home/palace/default.nix) — host
+  hardware, networkd (links/bridges/SR-IOV), LVM.
+- [`nixos/boxes/home/palace/vms/default.nix`](../../../nixos/boxes/home/palace/vms/default.nix) —
+  VM instances and boot ordering.
@@ -0,0 +1,81 @@
+# river
+
+The primary home router. A VM on `palace`, built from `routing-common` at index 0; it runs PPPoE
+to Digiweb directly on the ISP's VLAN 10 and normally holds the master side of the VRRP pair with
+[`stream`](stream.md).
+
+- **Source:** [`nixos/boxes/home/palace/vms/river.nix`](../../../nixos/boxes/home/palace/vms/river.nix)
+  (shared router config: [`routing-common`](../../../nixos/boxes/home/routing-common), index 0)
+- **Host:** VM on `palace`
+
+## Role
+
+- Primary of the router pair: `routing-common` index 0 → keepalived starts `MASTER` (priority
+  255), kea serves the lower-half DHCP pools, the zone's SOA/`ns1` point here.
+- Everything from `routing-common`: keepalived/VRRP ([`keepalived.nix`](../../../nixos/boxes/home/routing-common/keepalived.nix)),
+  PowerDNS recursor + authoritative with a blocklist ([`dns.nix`](../../../nixos/boxes/home/routing-common/dns.nix)),
+  kea DHCP + DDNS ([`kea.nix`](../../../nixos/boxes/home/routing-common/kea.nix)), radvd
+  ([`radvd.nix`](../../../nixos/boxes/home/routing-common/radvd.nix)), NAT/firewall, the
+  `as211024` L2 mesh link back to colony, `iperf3`, `nginx`. See
+  [networking.md](../../networking.md) for the logical view and [README.md](README.md#router-vips)
+  for the floating VIPs.
+- **Netboot server** for `sfh` and `castle` (`my.netboot.server`): iPXE/TFTP at
+  `boot.h.nul.ie` (a CNAME to `river-hi`), served from the `lo` address `192.168.72.1` and
+  restricted to the hi/lo prefixes.
+- **NVMe-oF client of `cellar`**: the VM's only local disk is an ESP (an installer ISO is also
+  still attached); `/nix` and `/persist` are LVs on the `nqn.2016-06.io.spdk:river` namespace
+  exported by `cellar` (`192.168.68.80`, RDMA).
+  The initrd brings up `lan-hi` with the RoCE modules (`roceBootModules`), and
+  `KeepConfiguration=static` on `lan-hi` stops networkd from dropping the NVMe-oF address on
+  reconfigure.
+- **SR-IOV VF passthrough**: the 100G `lan` NIC is VF 1 of palace's `et100g` (MAC
+  `52:54:00:8a:8a:f2`, MTU 9000). All router VLANs — hi/lo/untrusted plus both WAN VLANs — are
+  tagged on top of it (`55-lan`).
+- Also carries a macvtap passthrough of palace's 1G `et1g0`, renamed `wan-old` — the pre-100G WAN
+  path, kept around with no L3 config today.
+- Deploy address `192.168.68.1` (`my.deploy.node.hostname`).
+
+## Network assignments
+
+<!-- assignments: river -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| river | as211024 | `10.100.50.2/24` | `2a0e:97c0:4df:0:1::1/64 gw 2a0e:97c0:4df:0:2::1` | — |  |
+| river-core | core | `192.168.64.1/24` | — | h.nul.ie |  |
+| river-hi | hi | `192.168.68.1/22` | `2a0e:97c0:4d0:1::1/64` | h.nul.ie |  |
+| river-lo | lo | `192.168.72.1/21` | `2a0e:97c0:4d0:2::1/64` | h.nul.ie |  |
+| river-ut | untrusted | `192.168.80.1/24` | `2a0e:97c0:4d0:3::1/64` | h.nul.ie |  |
+<!-- assignments-end -->
+
+## WAN (Digiweb PPPoE)
+
+- `services.pppd` peer `digiweb` attaches PPPoE directly to `wan-pon-isp`
+  (`plugin pppoe.so wan-pon-isp`) — the **raw ISP VLAN 10** (`vlans.pon-isp`), trunked
+  untranslated from the ONT through `brian` and `dave` (the fabric side is in
+  [switches.md](switches.md)). The netdev is up with no L3; `MTUBytes=1508` (baby jumbo) absorbs
+  PPPoE's 8-byte overhead so the `wan` ppp interface gets a clean `mtu`/`mru 1500`.
+- The Digiweb static IP (`84.203.124.128`, `elemAt routersPubV4 0`) is requested as the local
+  address in IPCP. Auth is the across-all-customers shared `digiweb@nga.digiweb.ie` / `digiweb`
+  (deliberately not a secret); `persist`, `maxfail 0`, 1s LCP echoes. `usepeerdns` is absent on
+  purpose — Digiweb's resolvers are ignored in favour of the local recursor.
+- **ONT management**: `wan-pon-ont` (VLAN 140, PVID'd at the `brian` edge) holds
+  `192.168.100.100/24` so the ONT web UI at `192.168.100.1` is reachable — the `.100` mirrors
+  stream's modem-management convention.
+- **`wan-online.target`** is the shared "public WAN is up" gate declared by `routing-common`;
+  here it is driven by the pppd hooks (`DefaultDependencies=false`, so nothing pulls it in
+  early): `ip-up` installs `default dev wan scope link metric 100` and starts the target,
+  `ip-down` stops it and deletes the route. Consumers (`ipsec`, `ipv6-clear-default-route`)
+  attach with `wantedBy` + `partOf`, so they re-load on every WAN flap.
+- The `wan-ifb` ingress-shaping pieces from `routing-common` are inert on this box: the CAKE
+  config itself is stream's, and `networkd-dispatcher` (which installs the `tc` mirror) is
+  `mkForce false` here pending scheduling testing.
+
+## Notable config files
+
+- [`nixos/boxes/home/palace/vms/river.nix`](../../../nixos/boxes/home/palace/vms/river.nix) — box
+  config: pppd, WAN VLANs, `wan-online.target` hooks, netboot server, NVMe-oF boot.
+- [`nixos/boxes/home/routing-common/default.nix`](../../../nixos/boxes/home/routing-common/default.nix) —
+  shared router definition (assignments, firewall/NAT, `as211024`).
+- [`nixos/boxes/home/palace/vms/default.nix`](../../../nixos/boxes/home/palace/vms/default.nix) —
+  the VM definition on `palace` (VF 1, macvtap, ESP disk).
@@ -0,0 +1,63 @@
+# sfh
+
+"Services for home" — the NixOS container host for the home site. A VM on `palace` that netboots
+from `river` and runs its root off NVMe-oF from `cellar`.
+
+- **Source:** [`nixos/boxes/home/palace/vms/sfh/`](../../../nixos/boxes/home/palace/vms/sfh)
+  (`default.nix`, `containers/`)
+- **Host:** VM on `palace`
+
+## Role
+
+- Runs the home NixOS containers via `my.containers.instances` (systemd-nspawn); each container is
+  its own `nixos.systems.*` entry rendered through `my.asContainer`.
+- **Netboot client** (`my.netboot.client.enable`): the VM has no boot disk — its `netboot` NIC
+  (MAC `52:54:00:a5:7e:93`, on palace's `lan-lo` bridge, `bootindex=1`) is matched by a kea
+  client-class on [`river`](river.md) and iPXE-boots from `boot.h.nul.ie`.
+- **Root on NVMe-oF**: `my.nvme.boot` connects to `nqn.2016-06.io.spdk:sfh` at `192.168.68.80`
+  ([`cellar`](cellar.md), RDMA) from the initrd (`lan-hi` up + `roceBootModules`); `/nix` and
+  `/persist` are LVs on that volume. `KeepConfiguration=static` on `lan-hi` protects the
+  NVMe-oF address from networkd reconfigures.
+- **Frigate footage disk**: palace passes the `hdds/frigate` LVM LV through as a virtio disk;
+  sfh mounts it at `/mnt/frigate` (by label) and bind-mounts it into the `hass` container at
+  `/var/lib/frigate`.
+- USB: two host ports are passed to the VM (qemu flags) for the Zigbee coordinator and webcam used
+  by `hass`; the nspawn unit gets `DeviceAllow` for `char-ttyUSB` and `char-video4linux`.
+
+## Network assignments
+
+<!-- assignments: sfh -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| sfh | hi | `192.168.68.81/22 gw 192.168.71.254` | `2a0e:97c0:4d0:1::4:2/64` | h.nul.ie |  |
+<!-- assignments-end -->
+
+## Networking
+
+Four NICs, all MTU 9000 where jumbo-capable:
+
+- `lan-hi` — SR-IOV VF 2, the box's own `hi` assignment (`192.168.68.81`).
+- `lan-hi-ctrs` — SR-IOV VF 3, no L3: the MACVLAN parent for the containers' `hi` legs
+  (`host0` inside each container).
+- `lan-core-ctrs` / `lan-lo-ctrs` — virtio NICs (bridged to palace's `lan-core` / `lan-lo`), no
+  L3: MACVLAN parents for containers that need a `core` or `lo` leg.
+
+The per-container MACVLAN wiring lives in `systemd.nspawn.*.networkConfig` in
+[`sfh/default.nix`](../../../nixos/boxes/home/palace/vms/sfh/default.nix).
+
+## Containers
+
+| Container | Role | Page |
+|---|---|---|
+| `hass` | Home Assistant + Frigate + MQTT | [sfh/containers/hass.md](sfh/containers/hass.md) |
+| `unifi` | UniFi controller | [sfh/containers/unifi.md](sfh/containers/unifi.md) |
+
+## Notable config files
+
+- [`nixos/boxes/home/palace/vms/sfh/default.nix`](../../../nixos/boxes/home/palace/vms/sfh/default.nix) —
+  box config: netboot/NVMe-oF boot, container instances, MACVLAN plumbing, Frigate disk.
+- [`nixos/boxes/home/palace/vms/sfh/containers/`](../../../nixos/boxes/home/palace/vms/sfh/containers) —
+  the container system definitions.
+- [`nixos/boxes/home/palace/vms/default.nix`](../../../nixos/boxes/home/palace/vms/default.nix) —
+  the VM definition on `palace` (VFs, USB passthrough, netboot NIC, `hdds/frigate` disk).
@@ -0,0 +1,61 @@
+# hass
+
+Home automation container: Home Assistant plus its supporting services (MQTT, camera restreaming,
+Frigate NVR), running on [`sfh`](../../sfh.md).
+
+- **Source:** [`nixos/boxes/home/palace/vms/sfh/containers/hass.nix`](../../../../../nixos/boxes/home/palace/vms/sfh/containers/hass.nix)
+- **Host:** NixOS container on `sfh`
+
+## Role
+
+- **Home Assistant** (`services.home-assistant`) — declarative config (`configWritable = false`),
+  components `esphome`, `zha`, `denonavr`, `webostv`, `androidtv_remote`, `heos`, `mqtt`, `wled`,
+  `met`, `google_translate`; custom components `alarmo`, `frigate`, `west_wood_club`; a custom
+  Irish Rail sensor (Glenageary ↔ Dublin Connolly). A `hass-cli` wrapper is on the box, wired to
+  the local server with a token from `my.secrets`.
+- **mosquitto** — MQTT broker (anonymous local listener; port 1883 allowed, alongside HTTP).
+- **go2rtc** — restreams the Reolink living-room camera (RTSP from `reolink-living-room`, on the
+  `lo` leg) and the office USB webcam (`/dev/video0` via ffmpeg).
+- **Frigate** (`services.frigate`, `frigate.h.nul.ie` — the `frigate` alt name on the `hi`
+  assignment) — records both restreamed cameras with 1-day retention; detection is disabled.
+- External access is via `https://hass.nul.ie` through the `middleman` reverse proxy
+  (`trusted_proxies`); internally it's `hass-ctr.h.nul.ie`.
+- Not a deploy-rs target (`my.deploy.enable = false`) — it's rendered via `my.asContainer` and
+  started by `sfh`'s `my.containers.instances`.
+
+## Network assignments
+
+<!-- assignments: hass -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| hass-ctr (frigate) | hi | `192.168.68.103/22 gw 192.168.71.254` | `2a0e:97c0:4d0:1::5:3/64` | h.nul.ie |  |
+| hass-ctr-lo | lo | `192.168.72.103/21` | `2a0e:97c0:4d0:2::5:3/64` | h.nul.ie |  |
+<!-- assignments-end -->
+
+## Storage
+
+Frigate footage lives on a **separate HDD LV**: `palace` passes the `hdds/frigate` LVM LV to the
+`sfh` VM, sfh mounts it at `/mnt/frigate`, and the container bind-mounts it at `/var/lib/frigate`
+(read-write). This keeps recording churn off the NVMe-oF root.
+
+## Devices
+
+Passed through from `sfh` (USB host ports on `palace`):
+
+- Nabu Casa Connect ZBT-1 Zigbee coordinator → `/dev/ttyUSB0` (used by `zha`).
+- USB webcam → `/dev/video0` (go2rtc's `webcam_office` stream).
+- `/dev/bus/usb/001/002` (raw USB device node).
+
+## Networking
+
+MACVLAN legs created from `sfh`'s container NICs: `host0` on `lan-hi-ctrs` (the `hi` assignment,
+alt name `frigate`, default gateway via the VIP) and `lan-lo` on `lan-lo-ctrs` (the `lo`
+assignment, no gateway) — the `lo` leg reaches the IoT devices (the Reolink camera lives there).
+
+## Notable config files
+
+- [`nixos/boxes/home/palace/vms/sfh/containers/hass.nix`](../../../../../nixos/boxes/home/palace/vms/sfh/containers/hass.nix) —
+  container system: Home Assistant, Frigate, mosquitto, go2rtc.
+- [`nixos/boxes/home/palace/vms/sfh/default.nix`](../../../../../nixos/boxes/home/palace/vms/sfh/default.nix) —
+  the `sfh` side: bind mounts, MACVLAN wiring, `DeviceAllow`.
@@ -0,0 +1,47 @@
+# unifi
+
+The UniFi network controller, running as a container on [`sfh`](../../sfh.md). It manages the
+home UniFi switch `brian` (see [switches.md](../../switches.md)).
+
+- **Source:** [`nixos/boxes/home/palace/vms/sfh/containers/unifi.nix`](../../../../../nixos/boxes/home/palace/vms/sfh/containers/unifi.nix)
+- **Host:** NixOS container on `sfh`
+
+## Status
+
+**Currently enabled.** The container spent a while disabled — its import was commented out of
+[`containers/default.nix`](../../../../../nixos/boxes/home/palace/vms/sfh/containers/default.nix)
+while there was no UniFi gear to manage — and was re-enabled when the UniFi switch `brian` was
+added, gaining a `core` leg (`unifi-ctr-core`) at the same time so it can reach the switch on its
+management network. It is imported, listed in `sfh`'s `my.containers.instances`, and
+`services.unifi.enable = true`.
+
+## Role
+
+- **UniFi controller** (`services.unifi`, `pkgs.unifi` on `mongodb-7_0`, firewall open; TCP 8443
+  allowed).
+- Not a deploy-rs target (`my.deploy.enable = false`) — it's rendered via `my.asContainer` and
+  started by `sfh`'s `my.containers.instances`.
+
+## Network assignments
+
+<!-- assignments: unifi -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| unifi-ctr-core | core | `192.168.64.21/24` | — | h.nul.ie |  |
+| unifi-ctr | hi | `192.168.68.100/22 gw 192.168.71.254` | `2a0e:97c0:4d0:1::5:1/64` | h.nul.ie |  |
+<!-- assignments-end -->
+
+## Networking
+
+Two MACVLAN legs from `sfh`'s container NICs: `host0` on `lan-hi-ctrs` (the `hi` assignment —
+`unifi-ctr`, `192.168.68.100`, default gateway via the VIP) and `lan-core` on `lan-core-ctrs`
+(the `core` assignment — `unifi-ctr-core`, `192.168.64.21/24`, no gateway). The `core` leg is
+how the controller talks to the switches: `brian` lives at `192.168.64.13` on `core`.
+
+## Notable config files
+
+- [`nixos/boxes/home/palace/vms/sfh/containers/unifi.nix`](../../../../../nixos/boxes/home/palace/vms/sfh/containers/unifi.nix) —
+  container system: UniFi service, assignments.
+- [`nixos/boxes/home/palace/vms/sfh/default.nix`](../../../../../nixos/boxes/home/palace/vms/sfh/default.nix) —
+  the `sfh` side: container instance, MACVLAN wiring.
@@ -0,0 +1,82 @@
+# stream
+
+The secondary home router. A physical Intel box, built from `routing-common` at index 1; its WAN
+is a DHCP lease from the Virgin Media cable modem, and it is dual-homed to both switches.
+
+- **Source:** [`nixos/boxes/home/stream.nix`](../../../nixos/boxes/home/stream.nix) (shared router
+  config: [`routing-common`](../../../nixos/boxes/home/routing-common), index 1)
+- **Host:** physical
+
+## Role
+
+- Backup of the router pair: `routing-common` index 1 → keepalived starts `BACKUP` (priority
+  254), kea serves the upper-half DHCP pools, the zone's `ns2` points here. Takes over all VIPs
+  when [`river`](river.md) is down — see [README.md](README.md#router-vips).
+- Runs the same `routing-common` services as river: keepalived/VRRP, PowerDNS recursor +
+  authoritative, kea DHCP + DDNS, radvd, NAT/firewall, the `as211024` L2 mesh, `iperf3`, `nginx`.
+- Intel box (`kvm-intel`, `intel_iommu=on`, microcode updates).
+- `octoprint` and `mjpg-streamer` (3D-printer services) are defined but **disabled**
+  (`enable = false`).
+- `my.deploy.node.hostname` is currently commented out (it was `192.168.68.2`).
+
+## Network assignments
+
+<!-- assignments: stream -->
+<!-- assignments-start -->
+| Name | Assignment | IPv4 | IPv6 | Domain | Notes |
+|---|---|---|---|---|---|
+| stream | as211024 | `10.100.50.3/24` | `2a0e:97c0:4df:0:1::2/64 gw 2a0e:97c0:4df:0:2::1` | — |  |
+| stream-core | core | `192.168.64.2/24` | — | h.nul.ie |  |
+| stream-hi | hi | `192.168.68.2/22` | `2a0e:97c0:4d0:1::2/64` | h.nul.ie |  |
+| stream-lo | lo | `192.168.72.2/21` | `2a0e:97c0:4d0:2::2/64` | h.nul.ie |  |
+| stream-ut | untrusted | `192.168.80.2/24` | `2a0e:97c0:4d0:3::2/64` | h.nul.ie |  |
+<!-- assignments-end -->
+
+## WAN (Virgin Media DHCP)
+
+- `wan` is a renamed igc NIC (`00:f0:cb:ee:ca:dd`) towards the cable modem. The modem segment is
+  switch **VLAN 130** — tagging is handled on the switch (`jim`), so the box interface itself is
+  untagged. The fabric side is in [switches.md](switches.md).
+- `DHCP=ipv4` pulls the public lease; `dhcpV4Config.UseDNS=false` and the interface DNS points at
+  the local recursor. `IPv6AcceptRA=false` — this is an IPv4-only WAN (public IPv6 arrives over
+  the tunnel, not this link).
+- A **static modem-management address** (`192.168.0.100/24`, host `.100` of `prefixes.modem.v4`)
+  sits on `wan` alongside the DHCP lease so the modem's web UI stays reachable; it has no
+  gateway.
+- **`wan-online.target`** wiring: `wan-wait-online.service` (a oneshot) polls until the DHCP
+  default route exists, and the target `requires`/`after`s it (`wantedBy multi-user.target`).
+  The route — not networkd's wait-online — is the gate because the permanent static modem address
+  would otherwise report "online" before the public lease arrives, letting `ipsec` start
+  unoriented (`left=` is the public IP) and never connect.
+- **CAKE QoS**: egress is shaped at the `wan` root qdisc (`Bandwidth=48M`); ingress is redirected
+  by `tc` (`mirred`, installed by the `networkd-dispatcher` rule in `routing-common`) into the
+  `wan-ifb` IFB at `Bandwidth=490M` with the DOCSIS overhead preset.
+
+### Modem specifics (de-shared from `routing-common`)
+
+The modem's management subnet shares the `wan` interface, which `routing-common` itself knows
+nothing about — it declares two per-box options
+([`routing-common/default.nix`](../../../nixos/boxes/home/routing-common/default.nix)) that this
+box sets:
+
+- `my.homeRouter.dns.wanSkipBroadcasts = [ 192.168.0.255 ]` — skip the modem subnet when
+  auto-selecting the router's own `wan` A record for the zone's LUA record.
+- `my.homeRouter.firewall.untrustedRejectV4 = [ 192.168.0.0/24 ]` — reject untrusted clients from
+  reaching the modem subnet (needed only because it shares `wan`; WAN egress is otherwise
+  accepted).
+
+## Switching (STP)
+
+`stream` is dual-homed to both switches: `lan-jim` (igc) and `lan-dave` (mlx4_en), both MTU 9000,
+are enslaved to the `lan` bridge with `STP=true`. [`routing-common/mstpd.nix`](../../../nixos/boxes/home/routing-common/mstpd.nix)
+runs a patched `mstpd` and forces RSTP on `lan` once it's routable, so exactly one uplink carries
+traffic at a time. (The remaining NICs are renamed `et2`/`et5` and left unconfigured.)
+
+## Notable config files
+
+- [`nixos/boxes/home/stream.nix`](../../../nixos/boxes/home/stream.nix) — box config: DHCP WAN,
+  modem management, CAKE, `wan-online.target` gate, STP bridge.
+- [`nixos/boxes/home/routing-common/default.nix`](../../../nixos/boxes/home/routing-common/default.nix) —
+  shared router definition (index 1).
+- [`nixos/boxes/home/routing-common/mstpd.nix`](../../../nixos/boxes/home/routing-common/mstpd.nix) —
+  RSTP on the `lan` bridge.
@@ -0,0 +1,203 @@
+# Home switches
+
+Reference for the two MikroTik switches on the home network — **jim** and **dave** — plus the
+Ubiquiti switch **brian**, and how the home boxes and the Digiweb WAN hang off them. These switches
+are **not** managed by this flake; they are configured by hand (RouterOS on jim/dave, UniFi on
+brian). It covers the physical topology, the VLAN map, and how the Digiweb WAN reaches river.
+
+In short: the Digiweb ISP VLAN (10) is trunked straight through to river (which runs PPPoE on it),
+and the ONT's untagged management is PVID'd onto VLAN 140 at brian, its edge switch. VLAN 10 is
+carried untranslated because a single ONT makes it unique on the fabric — see
+[the WAN path](#the-digiweb-wan-path-trunked-vlan-10--pvid-140) and
+[why not translation](#why-not-translation-for-one-ont). The router side lives in
+[river.md](river.md); the logical network map in [networking.md](../../networking.md).
+
+## The switches
+
+| | jim | dave | brian |
+|---|---|---|---|
+| Identity | `jim-sw` | `dave-sw` | (UniFi) |
+| Model | CRS326-24G-2S+ | CRS504-4XQ | Ubiquiti 10GBASE-T |
+| Switch chip | Marvell 98DX3236 | Marvell 98DX4310 (+ Atheros 8227 for the 1G mgmt port) | — |
+| OS | RouterOS 7.18 | RouterOS 7.18 | UniFi |
+| Ports | 24×1G + 2×SFP+ | 4×QSFP28 (100G, breakout-capable) + 1G mgmt | 10GBASE-T |
+| Bridge | `main`, `vlan-filtering=yes` | `main`, `vlan-filtering=yes` | UniFi VLAN profiles |
+
+jim and dave run a single hardware-offloaded bridge (`main`) with VLAN filtering. Access to the
+MikroTiks is SSH as `admin` / `admin` by short hostname (see [Accessing the switches](#accessing-the-switches)).
+Only jim and dave can do hardware VLAN translation (`/interface ethernet switch rule` on the Marvell
+chips); brian cannot rewrite tags, only trunk/PVID them.
+
+## Physical topology
+
+The ONT terminates on brian; jim's `wan-pon-in` (`sfp-sfpplus2`) is a spare SFP+ port.
+
+```
+                                                  Virgin Media cable modem
+                                                            │
+                                                            │ (VLAN 130)
+                                                            │ wan1/wan2/wan-in
+       ┌──────────────────────────────────────────────────┴──┐
+       │  jim   CRS326-24G-2S+  (Marvell 98DX3236)             │
+       │  1G edge ports: castle, fort, pronter, laptop-dock,   │
+       │  palace-kvm, ups, ether15-20, wan1/wan2/wan-in        │
+       │  wan-pon-in (= sfp-sfpplus2)  ← spare SFP+            │
+       └───────┬───────────────────────────────────┬─────────┘
+   dave-uplink ┤ (= sfp-sfpplus1)                    │ stream, palace (1G secondaries)
+       10G trunk│                                     └── stream (box, dual-homed)
+       ┌─────────┴───────────────────────────────────────────────┐
+       │  dave   CRS504-4XQ  (Marvell 98DX4310)                    │
+       │  jim-downlink = qsfp28-3-1                                │
+       └──┬───────────┬──────────┬───────────────────┬────────────┘
+palace(100G)          │ castle   │ stream            │ brian-downlink
+= qsfp28-1-1          │          │                   │ (802.3ad LAG: brian1+brian2)
+     ┌────────────────┴─┐        …                   │
+     │ palace host      │              Digiweb (PPPoE) via ONT
+     │  └ river (VM)    │                       │
+     └──────────────────┘                 ┌─────┴─────┐
+   ↑ river's WAN + LAN ride the 100G link │    ONT    │ untagged mgmt 192.168.100.1
+                                          └─────┬─────┘ + tagged VLAN 10 (ISP)
+                                                │
+                                       ┌────────┴────────┐
+                                       │  brian (UniFi)  │  PVID 140 on the ONT port,
+                                       │  10GBASE-T      │  tagged VLAN 10 allowed through
+                                       └────────┬────────┘
+                                                └── brian-downlink LAG up to dave
+```
+
+Notes:
+- **river** runs as a VM on the **palace** host; its uplink is dave's 100G `palace` port. jim also
+  has 1G `palace`/`stream` ports, but those are secondary links and do **not** carry the WAN.
+- **stream** (the second router box) is dual-homed to both jim and dave (STP picks the active path).
+- **brian** is a Ubiquiti 10GBASE-T switch, downlinked from dave over an **802.3ad LAG**
+  (`brian-downlink` = `brian1` + `brian2`, layer-2 hash). It hosts the ONT.
+
+## VLANs
+
+| VLAN | Name | Purpose |
+|---|---|---|
+| — (native) | core | Switch management, `192.168.64.0/24` (jim `.10`, dave `.11`, brian `.13`) |
+| 100 | hi | High-performance / jumbo network (MTU 9000) |
+| 110 | lo | Standard LAN |
+| 120 | untrusted | Guest / untrusted network |
+| 130 | wan | **stream's WAN** — Virgin Media cable modem (untagged on jim's `wan1`/`wan2`/`wan-in`) |
+| 140 | wan-pon-ont | ONT management, `192.168.100.0/24` (PVID'd at the ONT edge) |
+| 10 | pon-isp | Digiweb ISP transport — **trunked straight through** to river, PPPoE runs on it |
+| 141 | wan-pon-isp | **Reserved** — the translated ISP VLAN for the future multi-ONT design |
+
+Switch L3 presence (`/interface vlan` on `main`) exists **only** for VLANs the switch is managed
+from — `hi` (100) and `lo` (110), plus native core. WAN and guest VLANs deliberately have no switch
+L3 interface.
+
+## The Digiweb WAN path (trunked VLAN 10 + PVID 140)
+
+The ONT presents two things on one wire:
+- **untagged** management traffic (`192.168.100.x`), and
+- **tagged VLAN 10** carrying the Digiweb ISP session (the BRAS requires VLAN 10).
+
+With a **single ONT** there's no reason to translate anything — VLAN 10 is unique on the fabric, so
+we just carry it end to end and let river run PPPoE directly on it:
+
+1. **Untagged mgmt → VLAN 140, at the ONT's edge switch (brian).** brian sets the ONT port's PVID to
+   140 so the untagged management traffic becomes VLAN 140, and allows tagged VLAN 10 through the
+   same port. river takes `192.168.100.100/24` on VLAN 140 (matching stream's modem-mgmt `.100`) to
+   reach the ONT web UI at `192.168.100.1`. Doing the PVID at the ONT-facing edge keeps it clean —
+   the untagged frames never share a domain with anything else.
+
+2. **VLAN 10 (ISP) trunked straight through, untranslated.** brian → dave → palace carry tagged
+   VLAN 10 by ordinary bridge-VLAN membership. No `/interface ethernet switch rule`, no pinning, no
+   asymmetric-learning issues — it's just a normal tagged VLAN. river attaches PPPoE to VLAN 10
+   directly (`wan-pon-isp` netdev = VLAN `pon-isp` = 10; baby-jumbo MTU 1508 so PPP nets a clean
+   1500).
+
+Net result: **river runs PPPoE single-tagged on VLAN 10 and holds a VLAN 140 address to reach the
+ONT.** See [`nixos/boxes/home/palace/vms/river.nix`](../../../nixos/boxes/home/palace/vms/river.nix)
+for the river side.
+
+```
+ONT ──(untagged + VLAN10)── brian ──(VLAN140 + VLAN10)── dave ──(VLAN140 + VLAN10)── river
+        ONT port             │ PVID140 + tagged 10        │ plain bridging
+                             └─ brian-downlink LAG ── dave ┘
+```
+
+### Why not translation (for one ONT)?
+
+Translation would swap VLAN 10 → 141 with two pinned hardware ACL rules to keep VLAN 10 off the rest
+of the fabric. That buys nothing with a single ONT — VLAN 10 is already unique, so trunking it is
+simpler and rule-free. Translation only earns its keep when **two** ONTs both deliver VLAN 10 and
+would collide (below).
+
+## Switch configuration
+
+How each switch is set up for the Digiweb WAN path. **Confirm any change on the box before applying**
+(see [Accessing the switches](#accessing-the-switches)).
+
+**brian (UniFi)** — hosts the ONT:
+- The ONT port has **native/untagged network = VLAN 140** (PVID) and is a **tagged member of VLAN 10**,
+  so the ONT's untagged management lands on 140 and its tagged ISP frames pass through.
+- The `brian-downlink` LAG up to dave trunks **tagged 140 + tagged 10** (alongside the LAN VLANs).
+
+**dave (RouterOS)** — trunks both WAN-pon VLANs to `brian-downlink` and `palace`. The ISP VLAN 10 row:
+```
+/interface bridge vlan add bridge=main vlan-ids=10 tagged=brian-downlink,palace
+```
+VLAN 140 also spans `brian-downlink,palace` (it carries a few other members too). No switch rules —
+this is plain tagged bridging.
+
+**jim (RouterOS)** — carries **none** of the Digiweb WAN path: no translation rules, and no VLAN
+10/140/141 rows. `wan-pon-in` (`sfp-sfpplus2`) sits at `pvid=1` as a spare port. jim only handles
+stream's VLAN-130 WAN and the LAN VLANs.
+
+## Future: multiple ONTs (per-port VLAN translation)
+
+If a second ONT arrives (e.g. a Digiweb line for stream, or a second river), trunking breaks: both
+ONTs deliver **tagged VLAN 10**, and plain bridge-VLAN filtering can't tell them apart. That's when
+translation earns its place — a switch rule matches on the **ingress port**, so each ONT's VLAN 10
+becomes a *distinct* fabric VLAN:
+
+- ONT-A port: VLAN 10 → **141** (→ river)
+- ONT-B port: VLAN 10 → **142** (→ stream / second river)
+- mgmt: PVID each ONT port onto its own VLAN (140, 143, …) so both ONTs' `192.168.100.1` stay in
+  separate L2/L3 domains.
+
+The forward direction isolates naturally (each ONT maps to a different fabric VLAN). The **return**
+direction is where port targeting is mandatory: both translate *back* to VLAN 10, so bridge VLAN 10
+now has two members and a plain FDB-miss flood would leak one ONT's upstream to the other. Each
+return must be pinned to its port with `new-dst-ports`:
+```
+# ONT-A: 141 in on palace → 10, forced out ONT-A's port
+# ONT-B: 142 in on stream → 10, forced out ONT-B's port
+```
+Each ONT port must also be a tagged member of bridge VLAN 10 for correct egress tagging (the missing
+piece that otherwise shows up as pppd "Timeout waiting for PADO"). The pins bypass the FDB, so the
+two ISP sessions never mix.
+
+**Why a new switch:** jim (the only box with spare SFP+ *and* the translation feature) has just
+**one** free SFP+ port, so it can't host two ONTs. The plan is a dedicated
+**CRS305-1G-4S+** (4×SFP+, same Marvell rule support) to land multiple ONTs and do the per-port
+translation there, feeding distinct fabric VLANs up to dave.
+
+## Accessing the switches
+
+The switches resolve by **short hostname** on the home network — the home routers serve their
+records in the home zone
+([`nixos/boxes/home/routing-common/dns.nix`](../../../nixos/boxes/home/routing-common/dns.nix):
+`jim` → hi `.10`, `dave` → hi `.11`, `brian` → core `.13`). From a box on the home network just
+`ssh admin@jim` / `admin@dave`. Non-interactive pattern (password auth, avoids the ssh-agent hang):
+
+```
+sshpass -p admin ssh -o IdentityAgent=none -o PubkeyAuthentication=no \
+  -o PreferredAuthentications=password -o StrictHostKeyChecking=accept-new \
+  -o UserKnownHostsFile=/tmp/sw_known_hosts admin@jim
+```
+
+**Always confirm config changes on the switch** (print the affected menu, apply, re-verify). brian
+is UniFi — configured through its controller, not RouterOS CLI.
+
+## Management IPs
+
+| | core (`192.168.64.0/24`) | hi (`192.168.68.0/22`) | lo (`192.168.72.0/21`) |
+|---|---|---|---|
+| jim | `.10` (on `main`) | `.10` | `.10` |
+| dave | `.11` (on `management`, the 1G Atheros port) | `.11` | `.11` |
+| brian | `.13` (core) | — | — |
@@ -8,7 +8,7 @@
          "ragenix",
          "nixpkgs"
        ],
-        "systems": "systems_7"
+        "systems": "systems_8"
      },
      "locked": {
        "lastModified": 1761656077,
@@ -67,6 +67,34 @@
        "type": "github"
      }
    },
+    "bun2nix": {
+      "inputs": {
+        "flake-parts": "flake-parts",
+        "nixpkgs": [
+          "pi-agent",
+          "nixpkgs"
+        ],
+        "systems": [
+          "pi-agent",
+          "systems"
+        ],
+        "treefmt-nix": "treefmt-nix_2"
+      },
+      "locked": {
+        "lastModified": 1778446047,
+        "narHash": "sha256-oQvcadh2BCkrog+SGrG6YffKJrveYpjj3TdQJWaKhaM=",
+        "owner": "nix-community",
+        "repo": "bun2nix",
+        "rev": "f2bc12af1a6369648aac41041ceeaa0b866599c6",
+        "type": "github"
+      },
+      "original": {
+        "owner": "nix-community",
+        "ref": "2.1.0",
+        "repo": "bun2nix",
+        "type": "github"
+      }
+    },
    "copyparty": {
      "inputs": {
        "flake-utils": "flake-utils_5",
@@ -256,6 +284,28 @@
        "type": "github"
      }
    },
+    "flake-parts": {
+      "inputs": {
+        "nixpkgs-lib": [
+          "pi-agent",
+          "bun2nix",
+          "nixpkgs"
+        ]
+      },
+      "locked": {
+        "lastModified": 1777988971,
+        "narHash": "sha256-qIoWPDs+0/8JecyYgE3gpKQxW/4bLW/gp45vow9ioCQ=",
+        "owner": "hercules-ci",
+        "repo": "flake-parts",
+        "rev": "0678d8986be1661af6bb555f3489f2fdfc31f6ff",
+        "type": "github"
+      },
+      "original": {
+        "owner": "hercules-ci",
+        "repo": "flake-parts",
+        "type": "github"
+      }
+    },
    "flake-utils": {
      "inputs": {
        "systems": "systems"
@@ -276,7 +326,7 @@
    },
    "flake-utils_10": {
      "inputs": {
-        "systems": "systems_9"
+        "systems": "systems_10"
      },
      "locked": {
        "lastModified": 1709126324,
@@ -294,7 +344,7 @@
    },
    "flake-utils_11": {
      "inputs": {
-        "systems": "systems_10"
+        "systems": "systems_11"
      },
      "locked": {
        "lastModified": 1705309234,
@@ -444,7 +494,7 @@
    },
    "flake-utils_9": {
      "inputs": {
-        "systems": "systems_8"
+        "systems": "systems_9"
      },
      "locked": {
        "lastModified": 1731533236,
@@ -586,6 +636,21 @@
        "type": "github"
      }
    },
+    "jail-nix": {
+      "locked": {
+        "lastModified": 1776230864,
+        "narHash": "sha256-YsEjjdOsGEzTeD+iT7ONh071BqWAOQWpzYVei3okAXE=",
+        "owner": "~alexdavid",
+        "repo": "jail.nix",
+        "rev": "404e7da9da5ab9aa643666682b2ba1312fa5fbe8",
+        "type": "sourcehut"
+      },
+      "original": {
+        "owner": "~alexdavid",
+        "repo": "jail.nix",
+        "type": "sourcehut"
+      }
+    },
    "libnetRepo": {
      "flake": false,
      "locked": {
@@ -795,6 +860,29 @@
        "type": "github"
      }
    },
+    "pi-agent": {
+      "inputs": {
+        "bun2nix": "bun2nix",
+        "jail-nix": "jail-nix",
+        "nixpkgs": [
+          "nixpkgs-unstable"
+        ],
+        "systems": "systems_7"
+      },
+      "locked": {
+        "lastModified": 1784984457,
+        "narHash": "sha256-y7jv+RTP0TkcJaDm7/D2xIFF9JDWKc6g5L3X3EUu3Eo=",
+        "owner": "lukasl-dev",
+        "repo": "pi.nix",
+        "rev": "fd2c62853f66a5803dae621769570f1165e59b80",
+        "type": "github"
+      },
+      "original": {
+        "owner": "lukasl-dev",
+        "repo": "pi.nix",
+        "type": "github"
+      }
+    },
    "pyproject-nix": {
      "inputs": {
        "nixpkgs": [
@@ -860,6 +948,7 @@
        "nixpkgs-mine-stable": "nixpkgs-mine-stable",
        "nixpkgs-stable": "nixpkgs-stable",
        "nixpkgs-unstable": "nixpkgs-unstable",
+        "pi-agent": "pi-agent",
        "ragenix": "ragenix",
        "sharry": "sharry"
      }
@@ -957,6 +1046,21 @@
        "type": "github"
      }
    },
+    "systems_11": {
+      "locked": {
+        "lastModified": 1681028828,
+        "narHash": "sha256-Vy1rq5AaRuLzOxct8nz4T6wlgyUR7zLU309k9mBC768=",
+        "owner": "nix-systems",
+        "repo": "default",
+        "rev": "da67096a3b9bf56a91d16901293e51ba5b49a27e",
+        "type": "github"
+      },
+      "original": {
+        "owner": "nix-systems",
+        "repo": "default",
+        "type": "github"
+      }
+    },
    "systems_2": {
      "locked": {
        "lastModified": 1681028828,
@@ -1098,6 +1202,28 @@
        "type": "github"
      }
    },
+    "treefmt-nix_2": {
+      "inputs": {
+        "nixpkgs": [
+          "pi-agent",
+          "bun2nix",
+          "nixpkgs"
+        ]
+      },
+      "locked": {
+        "lastModified": 1775636079,
+        "narHash": "sha256-pc20NRoMdiar8oPQceQT47UUZMBTiMdUuWrYu2obUP0=",
+        "owner": "numtide",
+        "repo": "treefmt-nix",
+        "rev": "790751ff7fd3801feeaf96d7dc416a8d581265ba",
+        "type": "github"
+      },
+      "original": {
+        "owner": "numtide",
+        "repo": "treefmt-nix",
+        "type": "github"
+      }
+    },
    "utils": {
      "inputs": {
        "systems": "systems_3"
@@ -49,6 +49,8 @@
    copyparty.inputs.nixpkgs.follows = "nixpkgs-unstable";
    hass-west-wood.url = "github:devplayer0/hass-west-wood";
    hass-west-wood.inputs.nixpkgs.follows = "nixpkgs-unstable";
+    pi-agent.url = "github:lukasl-dev/pi.nix";
+    pi-agent.inputs.nixpkgs.follows = "nixpkgs-unstable";
  };

  outputs =
@@ -8,6 +8,8 @@ let
  inherit (lib.my) mkOpt' dummyOption;
 in
 {
+  imports = [ inputs.pi-agent.homeModules.default ];
+
  options = with lib.types; {
    my = {
      isStandalone = mkOption {
@@ -27,6 +29,7 @@ in
      };
    };
  };
+
  config = mkMerge [
    {
      my = {
@@ -229,6 +232,7 @@ in
          inputs.deploy-rs.overlays.default
          inputs.boardie.overlays.default
          inputs.nixGL.overlays.default
+          inputs.pi-agent.overlays.default
        ];
        config = {
          allowUnfree = true;
@@ -71,8 +71,10 @@ in

            python3Packages.python-lsp-server
            nil # nix language server
+            nixd # another nix language server
            zls # zig language server
            rust-analyzer
+            pyright

            cowsay
            fortune
@@ -85,6 +87,17 @@ in

            ffmpeg-full
            xournalpp
+
+            (pkgs.symlinkJoin {
+              name = "pi-coding-agent";
+              buildInputs = [ pkgs.makeWrapper ];
+              paths = [ pkgs.pi-coding-agent-bun ];
+              postBuild = ''
+                wrapProgram $out/bin/pi \
+                  --set NPM_CONFIG_PREFIX ${config.home.homeDirectory}/.pi/npm/ \
+                  --prefix PATH : ${pkgs.lib.makeBinPath [ pkgs.nodejs_latest ]}
+              '';
+            })
          ];
        };

@@ -312,9 +312,16 @@ rec {
      lo = 110;
      untrusted = 120;
      wan = 130;
-      wan-pon = 131;

+      # Digiweb delivers the ISP VLAN (pon-isp, 10) single-tagged at the ONT alongside the ONT's
+      # own untagged management traffic. With a single ONT we trunk pon-isp (10) straight through
+      # the switches to river (PPPoE runs directly on it), and the switch at the ONT edge PVIDs the
+      # untagged management port onto wan-pon-ont (140). wan-pon-isp (141) is reserved for the
+      # future multi-ONT case, where per-port VLAN translation on a dedicated switch swaps each
+      # ONT's VLAN 10 to a distinct fabric VLAN (see docs/sites/home/switches.md).
      pon-isp = 10;
+      wan-pon-ont = 140;
+      wan-pon-isp = 141;
    };
    hiMTU = 9000;
    routers = [
@@ -330,6 +337,9 @@ rec {
      modem = {
        v4 = "192.168.0.0/24";
      };
+      ont = {
+        v4 = "192.168.100.0/24";
+      };
      all = {
        v4 = "192.168.64.0/18";
        v6 = "2a0e:97c0:4d0::/60";
@@ -9,16 +9,19 @@
    configuration = { lib, modulesPath, pkgs, config, assignments, allAssignments, ... }:
    let
      inherit (builtins) elemAt;
-      inherit (lib) mkForce mkMerge mkIf;
-      inherit (lib.my) networkdAssignment mkVLAN;
+      inherit (lib) mkForce mkMerge;
+      inherit (lib.my) net networkdAssignment mkVLAN;
      inherit (lib.my.c) networkd;
      inherit (lib.my.c.home) vlans domain prefixes roceBootModules routersPubV4;

-      # Digiweb currently delivers the ISP VLAN (pon-isp, 10) single-tagged, so PPPoE runs on a
-      # VLAN 10 sitting directly on the physical WAN link. Flip this to true to nest it back
-      # inside the wan-pon (131) transport VLAN — double-stacking also needs QinQ (tag-stacking)
-      # on the switch feeding the ONT, or the BRAS never answers PADI.
-      wanStacked = false;
+      # river reaches the ONT over its 100G `lan` uplink to the dave switch (which downlinks to
+      # brian, where the ONT lands). Digiweb delivers the ISP VLAN (pon-isp, 10) single-tagged at
+      # the ONT alongside the ONT's untagged management traffic. With a single ONT there's no VLAN
+      # collision, so the switches simply trunk the ISP's VLAN 10 straight through to river (PPPoE
+      # runs directly on it) and PVID the ONT's untagged management port onto wan-pon-ont (140).
+      # river takes .100 in the ONT's /24 (matching stream's modem-mgmt .100) to reach its web
+      # UI at 192.168.100.1. (See docs/sites/home/switches.md for the switch side and the multi-ONT plan.)
+      ontV4 = net.cidr.host 100 prefixes.ont.v4;

      # river is routing-common index 0; the Digiweb static IP we request via IPCP
      pubV4 = elemAt routersPubV4 0;
@@ -93,7 +96,7 @@
              enable = true;
              # Password is shared across all Digiweb customers, so no need for a secret
              config = ''
-                plugin pppoe.so wan-vlan-inner
+                plugin pppoe.so wan-pon-isp
                name "digiweb@nga.digiweb.ie"
                password "digiweb"
                # request our static IP as the local address in IPCP (local:remote, remote left open)
@@ -115,9 +118,10 @@
          };
        };

-        # PPPoE WAN (Digiweb): pppd owns the `wan` interface on top of VLAN 10, and its
-        # ip-up/ip-down hooks toggle the shared wan-online.target. Nothing else Wants the
-        # target, so it stays inactive until the link is actually up.
+        # PPPoE WAN (Digiweb): pppd owns the `wan` interface on top of wan-pon-isp (the switch's
+        # swap of the ISP's VLAN 10), and its ip-up/ip-down hooks toggle the shared
+        # wan-online.target. Nothing else Wants the target, so it stays inactive until the link
+        # is actually up.
        systemd.targets.wan-online.unitConfig.DefaultDependencies = false;

        environment.etc = {
@@ -143,8 +147,9 @@

        systemd.network = {
          netdevs = mkMerge [
-            (mkIf wanStacked (mkVLAN "wan-vlan-outer" vlans.wan-pon))
-            (mkVLAN "wan-vlan-inner" vlans.pon-isp)
+            (mkVLAN "wan-pon-ont" vlans.wan-pon-ont)
+            # The ISP VLAN is trunked through untranslated, so this is the raw pon-isp (10)
+            (mkVLAN "wan-pon-isp" vlans.pon-isp)
          ];

          links = {
@@ -176,24 +181,28 @@

          networks = {
            "55-lan" = {
-              # outer transport VLAN when stacked, otherwise the ISP VLAN directly on lan
-              vlan = [ (if wanStacked then "wan-vlan-outer" else "wan-vlan-inner") ];
+              # both WAN VLANs arrive single-tagged on the 100G uplink to dave: wan-pon-ont (140,
+              # the ONT's management, PVID-tagged at the brian edge) and the ISP's VLAN 10, trunked
+              # straight through
+              vlan = [ "wan-pon-ont" "wan-pon-isp" ];
            };
            # So we don't drop the IP we use to connect to NVMe-oF!
            "60-lan-hi".networkConfig.KeepConfiguration = "static";

-            "70-wan-vlan-outer" = mkIf wanStacked {
-              matchConfig.Name = "wan-vlan-outer";
-              vlan = [ "wan-vlan-inner" ];
-              networkConfig = networkd.noL3;
-              # baby jumbo: carries the inner VLAN's frames, whose 4B tag counts as payload
-              # at this layer, so it needs 1512 (inner's 1508B payload + the inner 802.1Q tag)
-              linkConfig.MTUBytes = "1512";
+            # ONT management: the brian edge PVIDs the ONT's untagged port onto wan-pon-ont, so give
+            # ourselves an address in its /24 to reach the ONT's web UI at 192.168.100.1.
+            "70-wan-pon-ont" = {
+              matchConfig.Name = "wan-pon-ont";
+              address = [ "${ontV4}/24" ];
+              linkConfig = {
+                RequiredForOnline = "no";
+                MTUBytes = "1500";
+              };
            };
-            # pppd attaches PPPoE to this; just needs to be up with no L3. Hangs off
-            # wan-vlan-outer when stacked, otherwise directly off lan (see "55-lan").
-            "71-wan-vlan-inner" = {
-              matchConfig.Name = "wan-vlan-inner";
+            # pppd attaches PPPoE to this; just needs to be up with no L3. This is the ISP's
+            # VLAN 10 trunked straight through from the ONT (no switch translation; see "55-lan").
+            "71-wan-pon-isp" = {
+              matchConfig.Name = "wan-pon-isp";
              linkConfig = {
                RequiredForOnline = "no";
                # baby jumbo: PPPoE's 8B overhead leaves a clean 1500 on ppp
@@ -11,12 +11,6 @@ in
 {
  nixos.systems."${name}" = {
    assignments = {
-      modem = {
-        ipv4 = {
-          address = net.cidr.host (254 - index) prefixes.modem.v4;
-          gateway = null;
-        };
-      };
      core = {
        name = "${name}-core";
        inherit domain;
@@ -100,9 +94,11 @@ in

    configuration = { lib, pkgs, config, assignments, allAssignments, ... }:
      let
-        inherit (lib) mkIf mkMerge mkForce;
-        inherit (lib.my) networkdAssignment;
+        inherit (lib) mkIf mkMerge mkForce optionalString concatStringsSep;
+        inherit (lib.my) mkOpt' networkdAssignment;
        inherit (lib.my.c) networkd;
+
+        cfg = config.my.homeRouter;
      in
      {
        imports = map (m: import m index) [
@@ -112,6 +108,20 @@ in
          ./kea.nix
        ];

+        # Per-box WAN-management specifics: the Virgin Media modem on stream lives on the `wan`
+        # interface itself, whereas river's ONT sits on its own interface. Declared as options the
+        # box sets so routing-common itself carries no modem/ONT knowledge.
+        options.my.homeRouter = with lib.types; {
+          dns.wanSkipBroadcasts = mkOpt' (listOf str) [ ] ''
+            Broadcast addresses to exclude when auto-selecting the router's own `wan` A record,
+            for extra static subnets that share the `wan` interface.
+          '';
+          firewall.untrustedRejectV4 = mkOpt' (listOf str) [ ] ''
+            IPv4 prefixes untrusted clients must be explicitly rejected from reaching. Only needed
+            for subnets sharing the `wan` interface, since `wan` egress is otherwise accepted.
+          '';
+        };
+
        config = {
          environment = {
            systemPackages = with pkgs; [
@@ -338,7 +348,7 @@ in
                    return
                  }
                  chain filter-untrusted {
-                    ip daddr ${prefixes.modem.v4} reject
+                    ${optionalString (cfg.firewall.untrustedRejectV4 != [ ]) "ip daddr { ${concatStringsSep ", " cfg.firewall.untrustedRejectV4} } reject"}
                    oifname wan accept
                    return
                  }
@@ -206,7 +206,7 @@ in
            ${name} IN LUA ${lib.my.dns.ifaceA {
              inherit pkgs;
              iface = "wan";
-              skipBroadcasts = [ (lib.my.netBroadcast prefixes.modem.v4) ];
+              skipBroadcasts = config.my.homeRouter.dns.wanSkipBroadcasts;
            }}
            ${otherName} IN LUA ${lib.my.dns.lookupIP {
              inherit pkgs;
@@ -6,11 +6,16 @@
    nixpkgs = "mine";
    home-manager = "mine";

-    configuration = { lib, pkgs, config, assignments, ... }:
+    configuration = { lib, pkgs, config, ... }:
    let
      inherit (lib) mkMerge;
-      inherit (lib.my) networkdAssignment;
+      inherit (lib.my) net;
      inherit (lib.my.c) networkd;
+      inherit (lib.my.c.home) prefixes;
+
+      # Static address on the Virgin Media modem's management subnet. Kept as a plain interface
+      # address (not a network assignment) since it's local to this box's WAN uplink.
+      modemV4 = net.cidr.host 100 prefixes.modem.v4;
    in
    {
      imports = [ ./routing-common/mstpd.nix ];
@@ -77,7 +82,7 @@
          };
        };

-        # wan carries a permanent static modem-management address (assignments.modem)
+        # wan carries a permanent static modem-management address (modemV4)
        # alongside the DHCP public IP, so wait-online@wan reports "online" as soon as
        # the static address is up - before the DHCP lease arrives. ipsec's left= is the
        # public IP, so gating on wait-online lets it start unoriented and never connect.
@@ -190,26 +195,28 @@
                CompensationMode=none
              '';
            };
-            "50-wan" = mkMerge [
-              (networkdAssignment "wan" assignments.modem)
-              {
-                matchConfig.Name = "wan";
-                DHCP = "ipv4";
-                dns = [ "127.0.0.1" "::1" ];
-                dhcpV4Config.UseDNS = false;
+            "50-wan" = {
+              matchConfig.Name = "wan";
+              # Static modem-management address alongside the DHCP public lease. It has no
+              # gateway, so the wan-wait-online gate keys off the DHCP default route instead.
+              address = [ "${modemV4}/24" ];
+              DHCP = "ipv4";
+              dns = [ "127.0.0.1" "::1" ];
+              dhcpV4Config.UseDNS = false;
+              # IPv4-only WAN (public IPv6 arrives over the tunnel, not this link).
+              networkConfig.IPv6AcceptRA = false;

-                qdiscConfig = {
-                  Parent = "ingress";
-                  Handle = "0xffff";
-                };
-                extraConfig = ''
-                  [CAKE]
-                  Parent=root
-                  Bandwidth=48M
-                  RTTSec=50ms
-                '';
-              }
-            ];
+              qdiscConfig = {
+                Parent = "ingress";
+                Handle = "0xffff";
+              };
+              extraConfig = ''
+                [CAKE]
+                Parent=root
+                Bandwidth=48M
+                RTTSec=50ms
+              '';
+            };
          };
        };

@@ -218,6 +225,12 @@
            key = "ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIPYTB4zeAqotrEJ8M+AiGm/s9PFsWlAodz3hYSROGuDb";
          };
          server.enable = true;
+          # The modem's management subnet shares the `wan` interface: skip its address when
+          # picking our own wan A record, and reject untrusted clients from reaching it.
+          homeRouter = {
+            dns.wanSkipBroadcasts = [ (lib.my.netBroadcast prefixes.modem.v4) ];
+            firewall.untrustedRejectV4 = [ prefixes.modem.v4 ];
+          };
          # deploy.node.hostname = "192.168.68.2";
        };
      };
@@ -83,6 +83,7 @@ in
          inputs.boardie.overlays.default
          inputs.copyparty.overlays.default
          inputs.hass-west-wood.overlays.default
+          inputs.pi-agent.overlays.default
        ];
        config = {
          allowUnfree = true;
@@ -11,4 +11,8 @@ in
  chocolate-doom2xx = callPackage ./chocolate-doom2xx { };
  windowtolayer = callPackage ./windowtolayer.nix { };
  swaylock-plugin = callPackage ./swaylock-plugin.nix { };
+
+  update-docs-assignments = pkgs.writeShellScriptBin "update-docs-assignments" ''
+    exec ${pkgs.python3}/bin/python3 ${../ci/update-docs-assignments.py} "$@"
+  '';
 }
Author	SHA1	Message	Date
jackos1998	f89ad33c26	ci: Add docs assignment-table auto-updater Update docs assignments / update (push) Successful in 57s Details Regenerate the marked network-assignment tables in docs/ from nixos.allAssignments (nix run .#update-docs-assignments, registered in pkgs/default.nix), preserving hand-written Notes cells. The workflow runs on pushes to the docs branch and commits any table updates back; flip its trigger to master when the docs branch merges.	2026-07-26 19:16:43 +01:00
jackos1998	d51f2d62b6	docs: Document the deployment Add a top-level README mapping the boxes and a full docs/ tree: topic pages (architecture, networking, deployment), per-site box pages for colony and home with containers nested under their hosts, remote and mobile boxes, the installer, and the home switch fabric reference (folded in from home-switches.md, with AGENTS.md and code comments retargeted to its new home). Box pages carry marked assignment tables that CI regenerates from nixos.allAssignments. AGENTS.md points at the new docs and keeps its terse agent version of the mechanics, referring to the topic pages for depth.	2026-07-26 19:16:43 +01:00
jackos1998	9b582bec7b	home-manager/gui: Add pi-coding-agent CI / Check, build and cache nixfiles (push) Failing after 2m29s Details	2026-07-26 18:55:20 +01:00
jackos1998	8c4223af18	AGENTS.md: Document switches, `deploy --boot`, and memory policy CI / Check, build and cache nixfiles (push) Successful in 45m13s Details - Note `CLAUDE.md` is a symlink to `AGENTS.md` (edit the real file). - Policy: prefer recording durable guidance here over agent memory. - Add the "Home switches" section (`jim`/`dave`/`brian`, access, confirm-before-apply) pointing at `home-switches.md`. - Document `deploy --boot` for staging a config without live-switching. - Note the `ssh-agent` hang workaround for outbound SSH (`ssh-machine`/`deploy`). - Add a comment convention: comment where useful, not for trivial code, and describe general purpose rather than the one-off reason for a change. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-23 21:35:50 +01:00
jackos1998	b95992735e	nixos/home: Trunk Digiweb VLAN 10 to river instead of translating With a single ONT the ISP's VLAN 10 is unique on the fabric, so there's no need to swap it to a private fabric VLAN. Carry pon-isp (10) trunked straight through the switches to river and run PPPoE directly on it, dropping the old wan-pon-isp (141) translation. The ONT's untagged management is PVID'd onto wan-pon-ont (140) at its edge switch (brian). wan-pon-isp (141) is kept as a constant, reserved for the future multi-ONT case where per-port VLAN translation disambiguates two VLAN-10 ONTs. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-23 21:29:16 +01:00
jackos1998	0d1562d150	home-manager/gui: Add more LSPs CI / Check, build and cache nixfiles (push) Successful in 44m33s Details	2026-07-23 01:37:33 +01:00
jackos1998	49de78d47f	nixos/home: De-share Virgin Media modem specifics from routing-common The modem management subnet is stream-specific (it shares stream's `wan` interface), so it no longer belongs in the shared routing-common. Drop the `modem` network assignment and configure the address directly on stream's `wan` (modemV4). The firewall reject and DNS wan-A-record skipBroadcasts that existed only because of the modem move to per-box options `my.homeRouter.{firewall.untrustedRejectV4,dns.wanSkipBroadcasts}`, set solely by stream. river needs neither: its ONT sits on a separate interface already blocked by the interface-based forward rules. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-23 01:35:08 +01:00
jackos1998	832d0b5542	nixos/home: Rework river WAN to single-tag VLAN 140/141 Digiweb's ONT delivers its untagged management traffic (192.168.100.0/24) alongside the tagged ISP VLAN 10. The switch feeding river now tags the ONT's untagged port as wan-pon-ont (140) and swaps the ISP VLAN 10 to wan-pon-isp (141) on ingress, so both reach river single-tagged: river takes an address in the ONT's /24 on 140 to reach its web UI, and PPPoE runs directly on 141. Replaces the old QinQ-into-131 (wanStacked) transport; the wan-pon (131) VLAN constant is dropped and prefixes.ont added. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-23 01:35:08 +01:00
jackos1998	4d4c05ea70	Symlink CLAUDE.md to AGENTS.md CI / Check, build and cache nixfiles (push) Successful in 49m29s Details	2026-07-22 18:44:03 +01:00