box/host/setup-host.sh
dan-claude-bot b429f5b65c feat(setup-host): refuse a claimed subnet, and BOX_SUBNET to move off one (#80)
Run inside a box, setup-host built a nested boxnet claiming the exact
subnet and gateway of the guest's own uplink: the guest then held
10.88.0.1 as a LOCAL address while its default route pointed at 10.88.0.1
as its GATEWAY, carried two connected routes for the subnet, and suffered
intermittent, self-recovering egress blackouts (~24-36 s, roughly hourly)
that no host-side check could attribute. The flagship use case funnels
agents toward exactly this — working on box, in a box — so the guard, not
the operator, has to catch it.

setup-host now scans the target subnet BEFORE any mutation: the default
route's gateway inside it (the smoking gun — that is this machine's own
uplink), or any non-boxnet interface holding an address in it, refuses
with the way out named (BOX_SUBNET). A prior boxnet owning the subnet is
the one claimant that is NOT a collision — the legitimate re-run — and an
existing bridge on a DIFFERENT subnet refuses too: setup-host converges a
bridge, it never re-addresses one under live leases.

BOX_SUBNET=<a.b.c.0/24> (validated at the gate, alongside BOX_DNS) moves
the whole stack from one place:
  · the bridge address derives (ipv4.address=<gw>/24)
  · the ACL's gateway carve-out derives — and the ruleset is now converged
    via 'network acl edit', not created once: the #80 escape hatch (moving
    the bridge) left the old /32 behind, stranding box DNS inside the
    10.0.0.0/8 drop, which presents as a dead resolver, never as a stale
    ACL
  · box-firewall reads the gateway off the live bridge ('|| true' because
    under pipefail an absent bridge would kill the script instead of
    answering "no bridge yet")
  · the drill, multiuser rehearsal and migrate-host probes derive the
    prefix from 'incus network get boxnet ipv4.address' instead of
    hardcoding 10.88 — a BOX_SUBNET host must not fail its own rehearsals

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-18 19:46:58 +00:00

354 lines
17 KiB
Bash
Executable file

#!/usr/bin/env bash
# One-time host setup: install Incus, create the isolated network + ACL and
# the box-net profile. Idempotent. Ubuntu 24.04 / Debian 13.
set -euo pipefail
self="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)/$(basename "${BASH_SOURCE[0]}")"
here="$(dirname "$(dirname "$self")")"
# Byte-identical copy of bin/box's box_tier() — this script must know the
# tier before any install tree exists, and test/cli.sh diffs the two copies
# so they cannot drift.
box_tier() {
[ "$(id -u)" -eq 0 ] && { printf 'admin\n'; return; }
local groups; groups="$(id -nG 2>/dev/null | tr ' ' '\n')"
if printf '%s\n' "$groups" | grep -qx incus-admin; then printf 'admin\n'
elif printf '%s\n' "$groups" | grep -qx incus; then printf 'restricted\n'
else printf 'none\n'
fi
}
# A restricted (incus-group) user cannot build daemon-global state, and
# telling them to escalate would be wrong twice: the stack is the admin's to
# own, and if 'box new' works for them it already exists. Say so and succeed —
# this must sit BEFORE the sudo resolution below, which would otherwise bury
# the honest answer under a privilege error. Gated on the tier, not on
# 'command -v sudo': having the sudo binary is not the same as holding a grant.
if [ "$(id -u)" -ne 0 ] && [ "$(box_tier)" = restricted ]; then
echo "You are in the 'incus' group (restricted tier): you manage your own boxes," >&2
echo "but the host's daemon-global stack is built by an admin. It is already set" >&2
echo "up if 'box new' works. Nothing for you to do here." >&2
exit 0
fi
# How we reach root, decided once. 'sudo' cannot be hardcoded: at UID 0 it is
# unnecessary, and on a minimal root image it is not installed at all — this
# script died on 'sudo: command not found' before doing anything, which made
# install.sh's deliberate root path unusable on exactly the hosts it was for.
if [ "$(id -u)" -eq 0 ]; then
SUDO=""
elif command -v sudo >/dev/null 2>&1; then
SUDO="sudo"
else
echo "ERROR: host setup needs root and 'sudo' was not found." >&2
echo " re-run this as root: $self" >&2
exit 1
fi
# --- The subnet, and the refusal to build on one something already owns -----
# (#80.) The stack's subnet was hardcoded, and running setup-host INSIDE a box
# gave the guest a nested boxnet claiming the exact subnet and gateway of its
# own uplink: the guest then held its gateway's address as a LOCAL address,
# carried two connected routes for the subnet, and suffered intermittent,
# self-recovering egress blackouts nobody could attribute — the host looked
# clean the whole time. The flagship use case funnels agents toward doing
# exactly this (working on box, in a box), so the guard must refuse BEFORE
# any mutation, and name the way out (BOX_SUBNET).
# BOX_SUBNET must be a /24 with a zero host octet — a.b.c.0/24. Everything
# the stack derives (the bridge address, the gateway carve-out, the firewall)
# assumes that shape, and a garbage value must die HERE, never inside an
# incus create or an nft rule.
valid_subnet() {
local o a="" b="" c="" rest=""
case "$1" in *.0/24) ;; *) return 1 ;; esac
IFS=. read -r a b c rest <<<"${1%/24}"
[ "$rest" = 0 ] || return 1
for o in "$a" "$b" "$c"; do
case "$o" in ''|*[!0-9]*) return 1 ;; esac
[ "${#o}" -le 3 ] && [ "$o" -le 255 ] || return 1
done
}
# Who, other than box's own bridge, already owns an address inside $1?
# Prints the claimant and succeeds when the subnet is claimed by a FOREIGNER;
# stays silent and fails when it is free — or held only by boxnet, which is
# the legitimate re-run, converging a stack this script built before. The
# most telling claimant is the default route's gateway: if it sits inside the
# target subnet, this machine's own uplink lives there — i.e. this is almost
# certainly the inside of a box. Pure over `ip` output, so test/cli.sh can
# drive it against canned tables with a shim ip.
subnet_claimant() {
local pfx hit
pfx="${1%0/24}"
hit="$(ip -4 route show default 2>/dev/null | awk -v p="$pfx" '
{ gw = ""; dev = ""
for (i = 1; i < NF; i++) { if ($i == "via") gw = $(i+1); if ($i == "dev") dev = $(i+1) }
if (index(gw, p) == 1 && dev != "boxnet") {
print "this machine\047s own DEFAULT GATEWAY (" gw " via " dev ")"; exit } }')"
if [ -z "$hit" ]; then
hit="$(ip -4 -o addr show 2>/dev/null | awk -v p="$pfx" '
$2 != "boxnet" && index($4, p) == 1 { print "interface " $2 " (" $4 ")"; exit }')"
fi
[ -n "$hit" ] && printf '%s\n' "$hit"
}
BOX_SUBNET="${BOX_SUBNET:-10.88.0.0/24}"
if ! valid_subnet "$BOX_SUBNET"; then
echo "ERROR: BOX_SUBNET='$BOX_SUBNET' is not a sane subnet — the stack takes a" >&2
echo " /24 with a zero host octet, e.g. BOX_SUBNET=10.89.0.0/24" >&2
exit 1
fi
BOX_GW="${BOX_SUBNET%.0/24}.1"
if hit="$(subnet_claimant "$BOX_SUBNET")"; then
echo "ERROR: refusing to build boxnet on $BOX_SUBNET — that subnet is already" >&2
echo " claimed here by $hit." >&2
echo " If that is this machine's uplink, you are INSIDE a box: a nested" >&2
echo " stack on the guest's own subnet captures its gateway address and" >&2
echo " blackholes its egress, intermittently (issue #80)." >&2
echo " Nothing was changed. To build a nested stack anyway, pick a free" >&2
echo " subnet: BOX_SUBNET=10.89.0.0/24 box setup-host" >&2
exit 1
fi
# A bridge this script built before is the one claimant that is NOT a
# collision — but it must AGREE with the target: setup-host converges an
# existing bridge, it never re-addresses one (boxes hold leases on it).
# ('|| true': under pipefail, `ip … dev boxnet` on a fresh host — no such
# device — would kill the script right here instead of answering "no bridge".)
have_gw="$(ip -4 -o addr show dev boxnet 2>/dev/null | awk '{ split($4, a, "/"); print a[1]; exit }' || true)"
if [ -n "$have_gw" ] && [ "$have_gw" != "$BOX_GW" ]; then
echo "ERROR: boxnet already exists on ${have_gw%.*}.0/24 and the target is $BOX_SUBNET" >&2
echo " setup-host converges an existing bridge, it never re-addresses one." >&2
echo " Re-run with the bridge's own subnet:" >&2
echo " BOX_SUBNET=${have_gw%.*}.0/24 box setup-host" >&2
echo " (or move the bridge first: incus network set boxnet ipv4.address $BOX_GW/24)" >&2
exit 1
fi
# apt, unattended-safe. install.sh now runs us without a human watching, and
# a fresh cloud image has apt-daily/unattended-upgrades holding the dpkg lock
# for the first minutes of its life — plain 'apt-get install' then waits on it
# in complete silence, indefinitely. Bound the wait and never prompt.
# 'env', not a bare VAR=val prefix: bash recognises assignments at PARSE time,
# so with $SUDO empty (we are root) 'DEBIAN_FRONTEND=x apt-get' would have
# already been parsed as a plain word and bash would try to EXECUTE it —
# 'DEBIAN_FRONTEND=noninteractive: command not found'. env is immune.
apt_get() {
$SUDO env DEBIAN_FRONTEND=noninteractive apt-get -o DPkg::Lock::Timeout=300 "$@"
}
if ! command -v incus >/dev/null; then
apt_get update
apt_get install -y incus
fi
if [ "$(id -u)" -eq 0 ]; then
# Root needs no group: UID 0 opens /var/lib/incus/unix.socket regardless of
# who owns it, and there is nothing to re-exec into. The HUMAN needs it — and
# under 'sudo install.sh' that is SUDO_USER, not the root we are running as.
# Adding root to incus-admin would be a no-op that also left the actual user
# locked out of their own boxes.
# NOTE: 'id -nG "$name"' here is deliberate and NOT the bug fixed below. That
# bug was asking the DATABASE about our own process; this asks the database
# about someone else's account, which is the only thing it can be asked.
login_user="${SUDO_USER:-}"
if [ -n "$login_user" ] && [ "$login_user" != root ]; then
if ! id -nG "$login_user" | grep -qw incus-admin; then
usermod -aG incus-admin "$login_user"
echo "added $login_user to incus-admin — log out and back in for your shell to pick it up"
fi
fi
# Group membership is a property of THIS PROCESS's credentials, not of the group
# database — and the two disagree for exactly as long as it matters here.
# 'id -nG "$USER"' names a user, so it reads /etc/group and reports incus-admin
# the instant usermod returns; the running shell's own credentials still lack
# it, because supplementary groups are fixed at login. So the old check passed
# on a same-session re-run, sailed into the incus calls below, and died on a
# permission error that named neither the group nor the re-login. Argless
# 'id -nG' asks the process what it actually holds, which is what incus checks
# when it opens /var/lib/incus/unix.socket.
elif ! id -nG | grep -qw incus-admin; then
$SUDO usermod -aG incus-admin "$USER"
# Then finish the job rather than adjourning it. Exiting 0 here was a
# success-shaped no-op: no boxnet, no ACL, no box-net profile, no firewall —
# and the burden of knowing that on the reader of a NOTE (#63). 'sg' runs us
# again with the new group in our credentials, no re-login, one invocation.
# The guard makes that at most one hop: if sg somehow lands without the
# group, we fail loudly instead of forking forever.
if [ -z "${BOX_SETUP_HOST_REEXEC:-}" ]; then
echo "added $USER to incus-admin — re-running under the new group (no re-login needed)"
export BOX_SETUP_HOST_REEXEC=1
exec sg incus-admin -c "$(printf '%q ' bash "$self" "$@")"
fi
echo "ERROR: still not in incus-admin after usermod + sg." >&2
echo " log out and back in, then re-run: box setup-host" >&2
exit 1
fi
# Storage pool + base config (safe to re-run: skipped once the pool exists).
# NOT 'incus admin init --minimal': minimal picks the 'dir' backend, which has
# no copy-on-write — every snapshot and clone is a FULL copy of the box's root,
# several GB and minutes apiece once a box is provisioned, against a workflow
# whose whole point is "log in once, snapshot, clone forever" (#29). btrfs on
# a loop device gives CoW (near-instant, near-free clones) with no
# partitioning. The preseed mirrors exactly what --minimal creates (pool,
# incusbr0, default profile) with only the driver deliberate; dir remains the
# fallback so a host that cannot do btrfs still works — just slowly, and it
# says so.
if ! incus storage show default >/dev/null 2>&1; then
driver=btrfs
command -v mkfs.btrfs >/dev/null 2>&1 || apt_get install -y btrfs-progs || driver=dir
if ! incus admin init --preseed <<PRESEED
storage_pools:
- name: default
driver: $driver
networks:
- name: incusbr0
type: bridge
profiles:
- name: default
devices:
root:
path: /
pool: default
type: disk
eth0:
name: eth0
network: incusbr0
type: nic
PRESEED
then
echo "storage: $driver preseed failed — falling back to --minimal (dir: every clone is a full disk copy)" >&2
incus admin init --minimal
fi
echo "storage: pool 'default' driver = $(incus storage show default | awk '/^driver:/ {print $2}')"
fi
# Isolated NAT network. IPv6 off: one less egress path to reason about.
# The default is 10.88 — not 10.87: a pre-rename host may still carry
# claudenet on 10.87 with legacy boxes attached — two bridges must not claim
# one subnet. BOX_SUBNET (validated and cleared by the #80 guard above) picks
# another /24; the gateway and every rule below derive from it.
incus network show boxnet >/dev/null 2>&1 || incus network create boxnet \
ipv4.address="$BOX_GW/24" ipv4.nat=true ipv6.address=none
# ACL: default egress allow (internet), explicit drops for private space.
# Gateway carve-out first so instance DNS (dnsmasq on the gateway) survives.
# 'edit' the full shipped ruleset, not create-once: the carve-out derives
# from BOX_SUBNET now, and a bridge moved off a colliding subnet (#80's
# escape hatch) left the OLD /32 behind — box DNS to the new gateway then
# died inside the 10.0.0.0/8 drop, looking like a dead resolver, not a stale
# ACL. A conditional 'rule add' cannot converge that (the stale carve-out
# would survive beside the new one); replacing the ruleset does, idempotently.
incus network acl show box-isolate >/dev/null 2>&1 || incus network acl create box-isolate
incus network acl edit box-isolate <<ACL
description: ""
egress:
- action: allow
destination: $BOX_GW/32
state: enabled
- action: drop
destination: 10.0.0.0/8
state: enabled
- action: drop
destination: 172.16.0.0/12
state: enabled
- action: drop
destination: 192.168.0.0/16
state: enabled
- action: drop
destination: 169.254.0.0/16
state: enabled
- action: drop
destination: 100.64.0.0/10
state: enabled
ingress: []
ACL
incus network set boxnet security.acls=box-isolate \
security.acls.default.egress.action=allow \
security.acls.default.ingress.action=drop
# A box must not be able to ENUMERATE its siblings, either. dnsmasq on the
# gateway serves DNS (that carve-out is what makes egress resolution work) and
# it holds a record for every instance on the network — so 'getent hosts <box>'
# from inside one box resolved another's name and address. Connection blocked,
# reconnaissance wide open. dns.mode=none stops it registering instance records;
# forwarding for public names is unaffected (verified live).
incus network set boxnet dns.mode=none
# A box's resolver must not be a function of the host's VPN posture (#33).
# The bridge's dnsmasq forwards to whatever sits in the HOST's /etc/resolv.conf
# at that moment. On a Tailscale/VPN host that is MagicDNS: box DNS flaps with
# the tailnet (this is what killed cold mints), and tailnet peer names and
# split-DNS zones RESOLVE from inside a box — name-level reconnaissance of a
# private network, the same shape as the sibling enumeration closed above.
# no-resolv detaches dnsmasq from the host's resolver entirely; server= pins a
# stable public upstream (override: BOX_DNS="ip ip…"). raw.dnsmasq is the
# lever — the bridge has no first-class upstream key. Verified live on the
# drill host: pin applied, box resolves, cold mint survives.
BOX_DNS="${BOX_DNS:-1.1.1.1 8.8.8.8}"
incus network set boxnet raw.dnsmasq \
"$(printf 'no-resolv\n'; for s in $BOX_DNS; do printf 'server=%s\n' "$s"; done)"
# Sibling isolation itself is NOT an ACL rule — an L3 ACL never sees frames
# switched between two ports of one bridge. It lives in box-firewall.sh
# as an nftables bridge-family rule. See the comment there; it is the reason
# boxes cannot reach each other.
# IPv6 stays off (ipv6.address=none, above). Every rule in the ACL and every
# rule in the firewall is IPv4-only, so IPv6 would be an uncovered path, not a
# feature. That is a contract, not a default.
# --- Firewall coexistence ---------------------------------------------------
# Hosts running UFW (INPUT drop) and/or Docker (FORWARD drop) silently eat
# boxnet traffic. Punch minimal, ordered holes; the Incus ACL still layers
# on top. The trailing deny also blocks instance -> host's own (public) IPs,
# which the RFC1918-only ACL cannot express. Rules live in
# box-firewall.sh; a boot-time systemd unit re-applies the runtime-only
# parts (nft table, DOCKER-USER) after every reboot.
# The no-UFW path drives nft directly, and a stock Debian 13 cloud image ships
# neither nftables nor UFW — install the dependency we are about to use.
if ! command -v ufw >/dev/null 2>&1 && ! command -v nft >/dev/null 2>&1; then
apt_get install -y nftables
fi
$SUDO install -m 755 "$here/host/box-firewall.sh" /usr/local/sbin/box-firewall
$SUDO install -m 644 "$here/host/box-firewall.service" /etc/systemd/system/
$SUDO systemctl daemon-reload
$SUDO systemctl enable box-firewall.service
# RESTART, not 'enable --now'. The unit is RemainAfterExit, so once it has run
# it stays "active" forever — and 'enable --now' does nothing to an active unit.
# Re-running setup-host after upgrading the tool therefore installed the new
# rules to /usr/local/sbin and never applied them: the host kept the old
# firewall, silently, and the box→box hole stayed open through a release that
# claimed to close it. Restart re-runs the script, which is idempotent by design.
$SUDO systemctl restart box-firewall.service
# incus-user is what serves the restricted tier (box grant). Debian 13 and
# Ubuntu 24.04 ship it inside the incus package; enabling it here makes the
# host tier-ready, and costs a host that never grants anyone nothing. Failure
# is a NOTE, not an error: the admin tier does not depend on it.
$SUDO systemctl enable --now incus-user.socket 2>/dev/null \
|| echo "NOTE: could not enable incus-user.socket — 'box grant' (the restricted tier) needs it; this Incus may not ship incus-user (#74)." >&2
# Profile — box-net, the placement contract: the isolated NIC and the root
# disk, nothing a template controls (resources are stamped per-instance from
# the template at mint time). A legacy claude-dev profile is left alone:
# Incus refuses to delete an in-use profile, and pre-rename boxes reference
# it until their last one is gone — teardown-host removes it then.
if ! incus profile show box-net >/dev/null 2>&1; then
incus profile create box-net
fi
incus profile edit box-net < "$here/profiles/box-net.yaml"
# The sibling drop is the one rule whose absence is invisible: everything keeps
# working, and boxes can simply reach each other. Assert it landed.
if $SUDO nft list table bridge box >/dev/null 2>&1; then
echo "Isolation: box-to-box drop is live (nft bridge table 'box')."
else
echo "WARNING: the box-to-box drop is NOT active — boxes can reach each other." >&2
echo " check: sudo /usr/local/sbin/box-firewall ; sudo nft list table bridge box" >&2
fi
echo "Host ready. Launch with: box new --name <box>"