README.md

@@ -20,7 +20,7 @@ PATH (`/usr/local/bin` when root). Re-run any time to upgrade.
 
 ## Commands
 
-### `rig bootstrap <control-plane|workload|runner>`
+### `rig bootstrap <control-plane|workload|runner|dev>`
 
 Run as root on the fresh box (over SSH). Convergent — safe to re-run; a
 second run changes nothing.
@@ -29,12 +29,13 @@ second run changes nothing.
 rig bootstrap control-plane --hostname my-coolify-box
 rig bootstrap workload --hostname my-prod-box
 rig bootstrap runner --hostname my-ci-box
+rig bootstrap dev --hostname dev-server
 ```
 
 - `--hostname <name>` — tailnet hostname (default: the role name)
 - `--ts-tag <tag>` — tailnet tag to advertise (default: `tag:server`;
-  the `runner` role defaults to `tag:ci` instead, and **refuses**
+  the `runner` role defaults to `tag:ci` and the `dev` role to `tag:local`,
-  `tag:server` outright — see below)
+  and both **refuse** `tag:server` outright — see below)
 
 What it does: installs `curl ca-certificates unattended-upgrades` (and
 enables periodic unattended upgrades); writes an sshd hardening drop-in
@@ -76,6 +77,50 @@ code, and advertising your server tag would extend every grant your servers
 hold (SSH between them, say) to that code. The refusal turns the worst
 misconfiguration from a documentation warning into a hard error.
 
+### `rig bootstrap dev`
+
+The **Incus claudebox host** — the one machine class rig didn't make. Everything
+else (control planes, workloads, runners) came up rig-made and reproducible; the
+box that runs the claudeboxes was hand-built, so "every box is rig-made" had a
+hole exactly where an agent runs. `dev` closes it.
+
+```sh
+rig bootstrap dev --hostname dev-server
+```
+
+On top of the shared machinery (the `00-rig.conf` sshd drop-in **and** its
+`sshd -T` effective-config assert, hostname convergence, tailscale join), `dev`
+installs and initialises **Incus**: `incus admin init --auto` gives it a default
+storage pool, the `default` profile, and a managed bridge (`incusbr0`). Init runs
+**once** — a second `bootstrap dev` detects the existing pool + profile root disk
+and skips it, so the run is a true no-op — and rig asserts the *effective* Incus
+state (`incus profile device show default`, `incus network list`) rather than
+trusting `init`'s exit code, the same discipline that caught the sshd first-wins
+bug.
+
+Three hard constraints, each enforced rather than documented:
+
+- **`tag:local`, never `tag:server`.** The ACL grants `tag:server → :22`, so a
+  dev host wearing the server tag hands the control plane free SSH. `dev` defaults
+  `--ts-tag` to `tag:local` and **refuses `tag:server`** (exit 2) — the correct
+  tag is the *only* reachable outcome, not a flag the operator remembers. This
+  already bit us: both M900s came up `tag:server` and had to be retagged by hand.
+- **The guest claudeboxes never join the tailnet.** The **host** joins; the
+  **guests** do not. An agent-inhabited box with its own tailnet node is a
+  foothold into the control plane, so operator SSH into a claudebox goes *through*
+  the host (ProxyJump), never a tunnel of its own. rig joins the host and stops —
+  there is deliberately no "enrol the guests" step, and if one is ever added,
+  that convenience is the bug.
+- **No credentials on the host.** Claudeboxes are creds-free by design; the
+  operator adds their own interactively. rig installs, templates, and holds no
+  credential — here as everywhere.
+
+**The rehearsal must assert *effective* state, not files rig wrote.** The existing
+Incus rehearsal runs in a pristine Debian container with no cloud-init drop-in, so
+it is structurally blind to the sshd first-wins bug. A dev-role rehearsal asserts
+what actually resolved: `sshd -T`, `incus info`, and `tailscale status --json`
+showing `tag:local` — then a second `bootstrap dev` proving a clean no-op.
+
 ### `rig coolify install --version <pin>`
 
 Control-plane box only. Installs Coolify at exactly the pinned version with

bin/rig

@@ -8,11 +8,13 @@ usage() {
 usage: rig <command> [args]
 
 commands:
-  bootstrap <control-plane|workload|runner> [--hostname <name>] [--ts-tag <tag>]
+  bootstrap <control-plane|workload|runner|dev> [--hostname <name>] [--ts-tag <tag>]
       OS plumbing on a pristine Debian box: hardening, unattended-upgrades,
       tailscale join. Prompts for a single-use tailnet pre-auth key
       (TS_AUTHKEY env overrides the prompt). Run as root. Role runner
-      defaults to tag:ci and refuses tag:server.
+      defaults to tag:ci and refuses tag:server. Role dev (the Incus
+      claudebox host) defaults to tag:local, refuses tag:server, and also
+      installs + initialises Incus.
   coolify install --version <pin>
       Pinned Coolify install (AUTOUPDATE=false). Control-plane box only.
   coolify backup install [options]

commands/bootstrap.sh

@@ -9,13 +9,20 @@ die()  { printf 'rig-bootstrap: ERROR: %s\n' "$1" >&2; exit "${2:-1}"; }
 
 usage() {
   cat <<'EOF'
-usage: rig bootstrap <control-plane|workload|runner> [--hostname <name>] [--ts-tag <tag>]
+usage: rig bootstrap <control-plane|workload|runner|dev> [--hostname <name>] [--ts-tag <tag>]
 
   --hostname  system + tailnet hostname (default: the role name)
   --ts-tag    tailnet tag to advertise (default: tag:server;
               role runner defaults to tag:ci and refuses tag:server —
               a CI box executes repo-controlled code, and your server
-              tag's grants must never extend to it)
+              tag's grants must never extend to it; role dev defaults to
+              tag:local and likewise refuses tag:server — the server tag's
+              ACL grants :22, so a mis-tagged Incus host would hand the
+              control plane free SSH)
+
+Role dev also installs and initialises Incus (the claudebox host). The HOST
+joins the tailnet; the guest claudeboxes deliberately do NOT — an
+agent-inhabited box on the tailnet is a foothold into the control plane.
 
 Provide the single-use tailscale pre-auth key via the TS_AUTHKEY env var, or
 enter it at the interactive prompt. It is used once and never written to disk.
@@ -25,15 +32,21 @@ EOF
 # --- args (validated before the root check, so errors are testable) ---------
 ROLE="${1:-}"
 case "$ROLE" in
-  control-plane|workload|runner) shift ;;
+  control-plane|workload|runner|dev) shift ;;
   -h|--help) usage; exit 0 ;;
-  "") usage >&2; die "role required (control-plane|workload|runner)" 2 ;;
+  "") usage >&2; die "role required (control-plane|workload|runner|dev)" 2 ;;
-  *) die "unknown role: $ROLE (want control-plane|workload|runner)" 2 ;;
+  *) die "unknown role: $ROLE (want control-plane|workload|runner|dev)" 2 ;;
 esac
 
 TS_HOSTNAME="$ROLE"
 if [ "$ROLE" = "runner" ]; then
   TS_TAG="tag:ci"
+elif [ "$ROLE" = "dev" ]; then
+  # The Incus claudebox host. tag:server's ACL grants it :22, so a dev box
+  # carrying it hands the control plane free SSH — so dev advertises tag:local,
+  # never tag:server (refused below, not merely defaulted). This already bit us:
+  # both M900s came up tag:server and had to be retagged by hand.
+  TS_TAG="tag:local"
 else
   TS_TAG="tag:server"
 fi
@@ -54,6 +67,12 @@ done
 if [ "$ROLE" = "runner" ] && [ "$TS_TAG" = "tag:server" ]; then
   die "role runner must not advertise tag:server" 2
 fi
+# A dev box is the Incus claudebox host. tag:server's ACL grants it :22, so a
+# dev box wearing it hands the control plane free SSH — the exact bug that made
+# the M900s retag-by-hand jobs. Correct-tag-only is enforced, not documented.
+if [ "$ROLE" = "dev" ] && [ "$TS_TAG" = "tag:server" ]; then
+  die "role dev must not advertise tag:server" 2
+fi
 
 # --- guards ------------------------------------------------------------------
 [ "$(id -u)" -eq 0 ] || die "must run as root"
@@ -201,9 +220,59 @@ else
   tailscale up --authkey="$TS_AUTHKEY" --hostname="$TS_HOSTNAME" --advertise-tags="$TS_TAG"
 fi
 
+# --- incus (dev role only) ----------------------------------------------------
+# The Incus claudebox host is the one machine class rig didn't make — it was
+# hand-built, so "every box is rig-made, reproducibly" had a hole exactly where
+# an agent runs. This block closes it: install Incus, initialise it once.
+#
+# NOTE — the guest claudeboxes deliberately do NOT join the tailnet. Only the
+# HOST joined above; an agent-inhabited box with its own tailnet node is a
+# foothold into the control plane, so operator SSH into a claudebox goes through
+# the host (ProxyJump), never a tunnel of its own. rig joins the host and stops.
+# There is intentionally no code here to enrol the guests: if bootstrap dev ever
+# grows a "join the guests too" convenience, that convenience is the bug.
+#
+# No credentials, either: claudeboxes are creds-free by design and the operator
+# adds their own interactively. rig installs, templates and holds nothing secret.
+if [ "$ROLE" = "dev" ]; then
+  if ! command -v incus >/dev/null 2>&1; then
+    log "installing incus"
+    # Debian 13 packages incus directly; keep the noninteractive frontend the
+    # base package block set, so a prompt never wedges an unattended bootstrap.
+    apt-get install -y -qq incus
+  else
+    log "incus already installed"
+  fi
+
+  # Initialise ONCE. `incus admin init --auto` is NOT idempotent — a second run
+  # errors out ("storage pool already exists"), which would break convergence.
+  # Detect a prior init by the artefacts --auto leaves behind — a storage pool
+  # AND a root disk on the default profile — and skip re-init when both exist,
+  # so a second `bootstrap dev` is a true no-op.
+  if incus storage list -f csv 2>/dev/null | grep -q . \
+     && incus profile device show default 2>/dev/null | grep -q 'type: disk'; then
+    log "incus already initialised; skipping incus admin init"
+  else
+    log "initialising incus (default storage pool, default profile, managed bridge)"
+    incus admin init --auto
+  fi
+
+  # Assert the EFFECTIVE state, not `init`'s exit code — the repo's "assert what
+  # resolved, not the action" rule (the same discipline that caught the sshd
+  # first-wins bug). A green `init` that left no root disk or no managed network
+  # is a host that cannot launch a claudebox; die here rather than at first use.
+  incus profile device show default 2>/dev/null | grep -q 'type: disk' \
+    || die "incus init did not leave a root disk on the default profile — check 'incus profile show default'"
+  incus network list -f csv 2>/dev/null | grep -q '^incusbr0,' \
+    || die "incus init did not create the managed bridge incusbr0 — check 'incus network list'"
+  log "incus initialised and verified (default profile has a root disk; incusbr0 present)"
+fi
+
 log "done — role ${ROLE}, hostname ${TS_HOSTNAME}"
 if [ "$ROLE" = "control-plane" ]; then
   log "next: rig coolify install --version <pin>"
 elif [ "$ROLE" = "runner" ]; then
   log "next: rig runner install --repo <owner/repo> --version <pin>"
+elif [ "$ROLE" = "dev" ]; then
+  log "next: launch claudeboxes on this host (guests stay off the tailnet; reach them via ProxyJump through this host)"
 fi

test/cli.sh

@@ -37,9 +37,15 @@ check "bootstrap: unknown role exits 2"    2 "unknown role"   "$ROOT/commands/bo
 check "bootstrap: unknown flag exits 2"    2 "unknown flag"   "$ROOT/commands/bootstrap.sh" workload --nope
 check "bootstrap: hostname needs value"    2 "needs a value"  "$ROOT/commands/bootstrap.sh" workload --hostname
 check "bootstrap: runner refuses tag:server" 2 "must not advertise tag:server" "$ROOT/commands/bootstrap.sh" runner --ts-tag tag:server
+# dev is the Incus claudebox host: tag:server would grant it :22 via the ACL, so
+# it must refuse the server tag exactly as runner does (correct-tag-only, not a
+# flag to remember). The incus init + effective tag:local assertion need a real
+# host, so they live in the rehearsal, not here.
+check "bootstrap: dev refuses tag:server"    2 "must not advertise tag:server" "$ROOT/commands/bootstrap.sh" dev --ts-tag tag:server
 if [ "$(id -u)" -ne 0 ]; then
   check "bootstrap: refuses non-root"      1 "must run as root" env TS_AUTHKEY=x "$ROOT/commands/bootstrap.sh" workload
   check "bootstrap: runner role parses, refuses non-root" 1 "must run as root" env TS_AUTHKEY=x "$ROOT/commands/bootstrap.sh" runner
+  check "bootstrap: dev role parses, refuses non-root"    1 "must run as root" env TS_AUTHKEY=x "$ROOT/commands/bootstrap.sh" dev
 else
   echo "skip: bootstrap non-root refusals (running as root)"
 fi