test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
#!/usr/bin/env bash
# drill.sh — end-to-end drill for claudebox, against a real Incus.
#
# ⚠ DESTRUCTIVE, AND MEANT TO BE. Run it on a THROWAWAY host you can format.
# It installs Incus, rewrites the host's firewall rules, installs a systemd
# unit, and creates and deletes instances. Never run it on a machine you care
# about.
#
# bash drill/drill.sh # asks first
# bash drill/drill.sh --yes # no prompt (CI, or you've read it)
# bash drill/drill.sh --ref main # drill a different branch of claudebox
# bash drill/drill.sh --keep-boxes # leave the boxes up to poke at
#
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
# Four phases:
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
# A. Incus semantics — the assumptions claudebox is built on, probed directly.
# These were only ever verified against a stub.
# B. The claudebox surface — the whole CLI, end to end, including the boundary.
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
# C. Isolation baseline — does the trust boundary actually hold? (#15 section A)
# D. Hardening rehearsal — #16's proposed changes, applied live and re-probed
# (#15 section B). FAILs here are design vetoes, not code bugs.
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
#
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
# Exit 0 = every check passed. The summary ends with a block of audit answers
# to paste into heavy-duty/claudebox#15.
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
#
# The file is one long 'probe && ok "..." || no "..."'. ok/no always return 0, so
# the C-may-run-when-A-is-true trap SC2015 warns about cannot fire here.
# shellcheck disable=SC2015
2026-07-13 22:05:32 +00:00
#
# NOT -e: a failing check is data, not a crash. NOT pipefail: half the checks
# are 'refusal 2>&1 | grep -q text' where the refusal exits 1/2 BY DESIGN, and
# 'grep -q' SIGPIPEs the left side on early match — pipefail turned both into
# false FAILs on the first live run. The pipeline verdict must be grep's alone.
set -u
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
REPO = " ${ CLAUDEBOX_REPO :- heavy -duty/claudebox } "
REF = " ${ CLAUDEBOX_REF :- main } "
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
YES = 0; KEEP = 0
SELF = " $( readlink -f " $0 " ) "
while [ $# -gt 0 ] ; do
case " $1 " in
--yes| -y) YES = 1; shift ; ;
--keep-boxes) KEEP = 1; shift ; ;
--repo) REPO = " $2 " ; shift 2 ; ;
--ref) REF = " $2 " ; shift 2 ; ;
--in-group) shift; break ; ; # internal: see below
-h| --help) sed -n '2,18p' " $0 " | sed 's/^# \{0,1\}//' ; exit 0 ; ;
*) echo " drill: unknown option: $1 " >& 2; exit 2 ; ;
esac
done
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
pass = 0; fail = 0; findings = ( ) ; audit = ( )
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
ok( ) { printf ' \033[32mPASS\033[0m %s\n' " $* " ; pass = $(( pass + 1 )) ; }
no( ) { printf ' \033[31mFAIL\033[0m %s\n' " $* " ; fail = $(( fail + 1 )) ; findings += ( " FAIL: $* " ) ; }
note( ) { printf ' \033[33mNOTE\033[0m %s\n' " $* " ; findings += ( " NOTE: $* " ) ; }
inf( ) { printf ' %s\n' " $* " ; }
phase( ) { printf '\n\033[1m══ %s\033[0m\n' " $* " ; }
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
aud( ) { audit += ( " $* " ) ; } # an answer for the #15 audit
wait_box( ) { # poll until exec answers (the VM agent can take a while), ~2 min
local b = " $1 " _i
for _i in $( seq 1 60) ; do
claudebox exec " $b " -- true >/dev/null 2>& 1 && return 0
sleep 2
done
return 1
}
fix(drill): no in-box probe can hang the run again
Run 4 hung 10+ minutes at C4. The cause is a shell trap, not a network
one: $( ) waits for stdout to CLOSE, not for the command to exit, so a
grandchild inheriting the exec session's stdout holds the substitution
open forever — and 'timeout' does not save you, because it kills the
wrapper, not the process holding the pipe. 'claudebox exec' made it
likelier still: it becomes 'sudo -u claude -i', a login zsh with
oh-my-zsh, which is a lot of machinery to keep a fd alive.
Every in-box read now goes through in_box()/box_curl(): incus exec
directly, stdin pinned to /dev/null, output landed in a file rather than
a pipe, and 'timeout -k' to hard-kill. box_curl returns the curl exit
code as its finding — for A3 the exit code IS the answer (0 connected,
7 refused-so-it-arrived, 28 dropped). C4 also prints what it is probing
before it probes, so a stall is locatable instead of silent.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 22:55:43 +00:00
# Read from inside a box WITHOUT ever hanging the drill.
#
# Two traps, both hit for real:
# · 'claudebox exec' becomes 'sudo -u claude -i' — a LOGIN zsh (oh-my-zsh and
# all). Fine for a person, needless machinery for a probe.
# · $( ) waits for stdout to CLOSE, not for the command to exit. A grandchild
# inheriting the exec session's stdout keeps the substitution open forever,
# and 'timeout' does not save you: it kills the wrapper, not the holder of
# the pipe. Run 4 hung 10+ minutes on exactly this.
# So: talk to 'incus exec' directly, pin stdin to /dev/null, land the output in
# a file (never a pipe), and hard-kill on timeout.
in_box( ) {
local b = " $1 " ; shift
local out; out = " $( mktemp) "
timeout -k 5 20 incus exec " $b " -- " $@ " >" $out " 2>/dev/null </dev/null
local rc = $?
cat " $out " ; rm -f " $out "
return " $rc "
}
fix(drill): the NIC inside a VM is enp5s0, not eth0 — read by subnet instead
A3, the one probe the whole audit exists for, has never fired in six
runs. It was never the network: the profile names the DEVICE eth0, but
inside a VM guest predictable naming renames it enp5s0, so every address
lookup — first the '(eth0)' CSV match, then 'ip addr show dev eth0' —
was hunting an interface that does not exist. I fixed that symptom twice
without ever questioning the assumption underneath it.
Read the address from inside the box and select by SUBNET (10.87.x, what
claudenet hands out) rather than by interface name. docker0's 172.17.x
is the decoy; the NIC's name is the guest's business, not ours.
A3 also gains a guard it should have had from the start: if the peer and
the source hold the SAME address, refuse to probe. That is not
hypothetical — clones were inheriting their source's machine-id, hence
its DHCP lease, hence its address, so 'archive → peer' was archive
probing itself and would have reported a cheerful 'reachable' as an
isolation failure.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 23:32:22 +00:00
# The box's address ON CLAUDENET. Three ways to get this wrong, all of them hit:
# · 'incus list' name filters are NOT regexes ("^b$" silently matches nothing)
# · its CSV quotes a multi-address box across lines
# · and the interface is NOT called eth0. The PROFILE names the device eth0,
# but inside a VM guest predictable naming renames it enp5s0. Six runs of
# A3 "not probed" were this, not the network.
# So: read it from inside the box, and select by SUBNET (10.87.x, what claudenet
# hands out) rather than by interface name — docker0 (172.17.x) is the decoy,
# and the NIC's name is the guest's business, not ours.
claudenet_ip( ) {
2026-07-13 22:05:32 +00:00
local b = " $1 " ip _i
for _i in $( seq 1 15) ; do
fix(drill): the NIC inside a VM is enp5s0, not eth0 — read by subnet instead
A3, the one probe the whole audit exists for, has never fired in six
runs. It was never the network: the profile names the DEVICE eth0, but
inside a VM guest predictable naming renames it enp5s0, so every address
lookup — first the '(eth0)' CSV match, then 'ip addr show dev eth0' —
was hunting an interface that does not exist. I fixed that symptom twice
without ever questioning the assumption underneath it.
Read the address from inside the box and select by SUBNET (10.87.x, what
claudenet hands out) rather than by interface name. docker0's 172.17.x
is the decoy; the NIC's name is the guest's business, not ours.
A3 also gains a guard it should have had from the start: if the peer and
the source hold the SAME address, refuse to probe. That is not
hypothetical — clones were inheriting their source's machine-id, hence
its DHCP lease, hence its address, so 'archive → peer' was archive
probing itself and would have reported a cheerful 'reachable' as an
isolation failure.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 23:32:22 +00:00
ip = " $( in_box " $b " ip -4 -o addr show scope global \
| awk '{ for (i = 1; i < NF; i++) if ($i == "inet" && $(i+1) ~ /^10\.87\./) { split($(i+1), a, "/"); print a[1]; exit } }' ) "
2026-07-13 22:05:32 +00:00
[ -n " $ip " ] && { printf '%s\n' " $ip " ; return 0; }
sleep 2
done
return 1
}
fix(drill): read curl's message — the exit code cannot tell you what happened
Boxes are isolated. The kernel says so: 'isolated on' on both live bridge
ports. A box gets 100% packet loss pinging its sibling, and its TCP
connect burns the full timeout and dies with "Could not connect to
server" — not "Connection refused", which is what an arriving packet
gets, instantly.
The drill called that a FAIL. curl exit 7 is "failed to connect", and it
covers BOTH a refusal (a RST came back — the packet ARRIVED) and an
unreachable host (nothing came back — it was DROPPED). Opposite
conclusions, one exit code, and the drill mapped 7 → "arrived". So for
two runs after the isolation fix had landed and was working, the drill
reported a working boundary as a broken one.
The words distinguish what the number cannot. box_probe now returns
reachable | refused | dropped by reading the message, ping corroborates,
and every isolation probe (sibling, box→host, RFC1918, host→box) uses it.
RUNS.md gains trap 12. It is the same disease as all eleven before it:
trusting a proxy for the fact instead of the fact.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 02:33:02 +00:00
# The probe. Its verdict comes from curl's MESSAGE, never from its exit code.
#
# curl exit 7 is "failed to connect" — and it covers BOTH of these:
# · "Connection refused" → a RST came back. The packet ARRIVED. Reachable.
# · "Could not connect to server" / "No route to host" → nothing came back at
# all. The frame went nowhere. ISOLATED.
# Opposite conclusions, one exit code. The drill mapped 7 → "it arrived" and so
# reported a WORKING boundary as a broken one, run after run, while the kernel
# had 'isolated on' the bridge ports the whole time. A refusal is instant; an
# unreachable host burns the timeout. The words say which; the number cannot.
#
# Never hangs: incus exec directly (no login shell), stdin pinned, output landed
# in a file rather than a pipe, hard kill on timeout.
box_probe( ) { # box_probe <box> <url> [timeout] → reachable | refused | dropped
local b = " $1 " url = " $2 " t = " ${ 3 :- 5 } " out rc msg
out = " $( mktemp) "
fix(drill): no in-box probe can hang the run again
Run 4 hung 10+ minutes at C4. The cause is a shell trap, not a network
one: $( ) waits for stdout to CLOSE, not for the command to exit, so a
grandchild inheriting the exec session's stdout holds the substitution
open forever — and 'timeout' does not save you, because it kills the
wrapper, not the process holding the pipe. 'claudebox exec' made it
likelier still: it becomes 'sudo -u claude -i', a login zsh with
oh-my-zsh, which is a lot of machinery to keep a fd alive.
Every in-box read now goes through in_box()/box_curl(): incus exec
directly, stdin pinned to /dev/null, output landed in a file rather than
a pipe, and 'timeout -k' to hard-kill. box_curl returns the curl exit
code as its finding — for A3 the exit code IS the answer (0 connected,
7 refused-so-it-arrived, 28 dropped). C4 also prints what it is probing
before it probes, so a stall is locatable instead of silent.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 22:55:43 +00:00
timeout -k 5 $(( t + 15 )) incus exec " $b " -- curl -sS -m " $t " -o /dev/null " $url " \
fix(drill): read curl's message — the exit code cannot tell you what happened
Boxes are isolated. The kernel says so: 'isolated on' on both live bridge
ports. A box gets 100% packet loss pinging its sibling, and its TCP
connect burns the full timeout and dies with "Could not connect to
server" — not "Connection refused", which is what an arriving packet
gets, instantly.
The drill called that a FAIL. curl exit 7 is "failed to connect", and it
covers BOTH a refusal (a RST came back — the packet ARRIVED) and an
unreachable host (nothing came back — it was DROPPED). Opposite
conclusions, one exit code, and the drill mapped 7 → "arrived". So for
two runs after the isolation fix had landed and was working, the drill
reported a working boundary as a broken one.
The words distinguish what the number cannot. box_probe now returns
reachable | refused | dropped by reading the message, ping corroborates,
and every isolation probe (sibling, box→host, RFC1918, host→box) uses it.
RUNS.md gains trap 12. It is the same disease as all eleven before it:
trusting a proxy for the fact instead of the fact.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 02:33:02 +00:00
>/dev/null 2>" $out " </dev/null
rc = $?
msg = " $( cat " $out " ) " ; rm -f " $out "
if [ " $rc " -eq 0 ] ; then echo reachable; return ; fi
case " $msg " in
*"Connection refused" *) echo refused ; ; # it ARRIVED, and was rejected
*) echo dropped ; ; # nothing came back
esac
fix(drill): no in-box probe can hang the run again
Run 4 hung 10+ minutes at C4. The cause is a shell trap, not a network
one: $( ) waits for stdout to CLOSE, not for the command to exit, so a
grandchild inheriting the exec session's stdout holds the substitution
open forever — and 'timeout' does not save you, because it kills the
wrapper, not the process holding the pipe. 'claudebox exec' made it
likelier still: it becomes 'sudo -u claude -i', a login zsh with
oh-my-zsh, which is a lot of machinery to keep a fd alive.
Every in-box read now goes through in_box()/box_curl(): incus exec
directly, stdin pinned to /dev/null, output landed in a file rather than
a pipe, and 'timeout -k' to hard-kill. box_curl returns the curl exit
code as its finding — for A3 the exit code IS the answer (0 connected,
7 refused-so-it-arrived, 28 dropped). C4 also prints what it is probing
before it probes, so a stall is locatable instead of silent.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 22:55:43 +00:00
}
fix(drill): read curl's message — the exit code cannot tell you what happened
Boxes are isolated. The kernel says so: 'isolated on' on both live bridge
ports. A box gets 100% packet loss pinging its sibling, and its TCP
connect burns the full timeout and dies with "Could not connect to
server" — not "Connection refused", which is what an arriving packet
gets, instantly.
The drill called that a FAIL. curl exit 7 is "failed to connect", and it
covers BOTH a refusal (a RST came back — the packet ARRIVED) and an
unreachable host (nothing came back — it was DROPPED). Opposite
conclusions, one exit code, and the drill mapped 7 → "arrived". So for
two runs after the isolation fix had landed and was working, the drill
reported a working boundary as a broken one.
The words distinguish what the number cannot. box_probe now returns
reachable | refused | dropped by reading the message, ping corroborates,
and every isolation probe (sibling, box→host, RFC1918, host→box) uses it.
RUNS.md gains trap 12. It is the same disease as all eleven before it:
trusting a proxy for the fact instead of the fact.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 02:33:02 +00:00
box_pings( ) { # box_pings <box> <ip> → 0 if it answers ICMP
timeout -k 5 20 incus exec " $1 " -- ping -c1 -W2 " $2 " >/dev/null 2>& 1 </dev/null
2026-07-13 23:02:48 +00:00
}
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
# --- stage 1: consent, install, then re-enter inside the incus-admin group ---
if [ " ${ IN_GROUP :- 0 } " != 1 ] ; then
if [ " $YES " -ne 1 ] ; then
cat <<EOF
This will, ON THIS HOST ( $( hostname) ) :
· install Incus and a systemd unit
· create a network ( claudenet) , an ACL, and a profile
· rewrite firewall rules ( nft or UFW, and Docker' s DOCKER-USER chain)
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
· create and destroy instances named: drill, clone, archive, peer, payroll, cbprobe, cbcopy
· mutate the network and profile mid-run to rehearse the #16 hardening
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
Only do this on a machine you can format.
EOF
[ -t 0 ] || { echo "drill: no TTY to confirm on — pass --yes if you mean it." >& 2; exit 2; }
printf 'Continue? [y/N] '
read -r reply
case " $reply " in y| Y| yes) ; ; *) echo "stopped." ; exit 1 ; ; esac
fi
phase " Installing claudebox ( $REPO @ $REF ) "
CLAUDEBOX_REPO = " $REPO " CLAUDEBOX_REF = " $REF " \
bash -c " $( curl -fsSL " https://raw.githubusercontent.com/ $REPO / $REF /install.sh " ) " \
|| { echo "install failed" ; exit 1; }
export PATH = " $HOME /.local/bin: $PATH "
phase "Host setup (Incus, claudenet, ACL, profile, firewall)"
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
# setup-host.sh installs nftables itself when neither nft nor UFW exists
2026-07-13 23:05:34 +00:00
# (a stock Debian 13 cloud image ships neither). This guard is a tripwire:
# if it fires, that fix regressed.
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
if ! command -v nft >/dev/null 2>& 1 && ! command -v ufw >/dev/null 2>& 1; then
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
note "neither nft nor ufw present pre-setup — setup-host.sh must install nftables itself (it fixed this once; watch that it still does)"
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
fi
2026-07-13 23:05:34 +00:00
# Sudo, up front and out loud. Later calls run unattended, and a password
# prompt swallowed by a '-qq' redirect looks exactly like a hang.
sudo -v || { echo "drill: need sudo (the host setup installs packages and firewall rules)" ; exit 1; }
# apt's lock is held by apt-daily / unattended-upgrades on a fresh cloud
# image, and 'apt-get -qq >/dev/null' waits for it in COMPLETE SILENCE —
# which is how run 5 looked stuck for minutes right after this header.
# Say what we are waiting for, and give up rather than hang forever.
if ! command -v incus >/dev/null 2>& 1; then
inf "installing incus (waiting for the apt lock if a background upgrade holds it)…"
if ! sudo DEBIAN_FRONTEND = noninteractive timeout 600 \
apt-get -o DPkg::Lock::Timeout= 300 install -y incus; then
echo "drill: 'apt-get install incus' failed or timed out." >& 2
echo " a background apt job usually holds the lock. check with:" >& 2
echo " sudo fuser -v /var/lib/dpkg/lock-frontend" >& 2
echo " systemctl status unattended-upgrades apt-daily.service" >& 2
exit 1
fi
else
inf "incus already installed — skipping apt"
fi
inf "running setup-host.sh (first pass: may only add you to incus-admin)…"
~/.local/share/claudebox/host/setup-host.sh || true
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
# The group we were just added to isn't in this shell's credentials yet.
2026-07-13 23:05:34 +00:00
inf "re-entering inside the incus-admin group…"
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
exec sg incus-admin -c " IN_GROUP=1 CLAUDEBOX_REPO=' $REPO ' CLAUDEBOX_REF=' $REF ' KEEP= $KEEP bash ' $SELF ' --in-group "
fi
export PATH = " $HOME /.local/bin: $PATH "
KEEP = " ${ KEEP :- 0 } "
fix(drill): clean the host BEFORE setup-host, and bound it
Run 6 stalled in setup-host.sh, which should be seconds on a host that
already has incus. The ordering was wrong: cleanup ran AFTER setup, so
setup-host reconfigured claudenet's ACLs while an aborted run's boxes
were still attached to that network — 'incus network set' then has to
push the change onto every live NIC. The same aborted run also left the
D-phase mutations (dns.mode=none, NIC filtering) in place, so setup was
converging against a moving target.
Boxes are now deleted and the mutations reverted first, so setup-host is
the no-op it should be. It is also bounded (5 min) and, on timeout,
prints the three things worth checking instead of hanging: instances
still on the network, the firewall unit, the incus daemon. Every cleanup
call gets its own timeout, so a wedged instance cannot stall the run
before the drill has printed a single line.
RUNS.md gains this as trap 8.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 23:09:51 +00:00
# CLEAN BEFORE SETUP, not after. setup-host.sh reconfigures the network's ACLs,
# and a previous run's boxes are still ATTACHED to that network — 'incus network
# set' then has to push the change onto every live NIC, which is how run 6
# stalled. An aborted run also leaves the D-phase mutations (dns.mode=none, NIC
# filtering) in place, so setup would be converging against a moving target.
# Take the boxes down and revert the mutations FIRST; then the host is a
# clean-ish slate and setup-host is the no-op it should be.
fix(drill): stop poisoning the host, and add a doctor to prove it
Run 8's cold mint failed with 'cloud-init status: error' — the box could
not resolve deb.debian.org, or claude.ai, or anything. The cause was not
in that run at all: run 7's phase D set dns.mode=none on claudenet, the
run ended before reverting it, and every box minted afterwards came up
with no DNS.
This is the worst failure mode the drill has: a poisoned host does not
fail the next run honestly, it produces confident wrong answers. It is
how a false design veto against #16 got posted, and it wasted a cold
mint plus an hour of diagnosis that had nothing to do with the code
under test.
Three defences:
· the phase-D revert is armed with a trap BEFORE the first mutation,
so it fires on any exit, Ctrl-C included;
· the revert is VERIFIED rather than fired into /dev/null, so a failed
unset can no longer masquerade as a successful one;
· the drill refuses to start on a host still carrying the mutations.
And drill/doctor.sh answers the question that kept being answered by
hand: what state is this host actually in? Network, profile, ACL,
leftover boxes, and whether a box can still resolve DNS — with --fix to
revert the leftovers.
RUNS.md gains trap 10.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 00:32:03 +00:00
# A host still carrying a previous run's phase-D mutations mints boxes with no
# DNS, and then reports the resulting breakage as a finding. Refuse to run.
2026-07-14 01:40:02 +00:00
# NOTE: dns.mode=none is now part of the SHIPPED stack (it closes the sibling
# DNS-enumeration leak), so it is no longer "dirt" from a rehearsal — do not
# revert it. Only the vetoed NIC filtering counts as leftover.
fix(drill): stop poisoning the host, and add a doctor to prove it
Run 8's cold mint failed with 'cloud-init status: error' — the box could
not resolve deb.debian.org, or claude.ai, or anything. The cause was not
in that run at all: run 7's phase D set dns.mode=none on claudenet, the
run ended before reverting it, and every box minted afterwards came up
with no DNS.
This is the worst failure mode the drill has: a poisoned host does not
fail the next run honestly, it produces confident wrong answers. It is
how a false design veto against #16 got posted, and it wasted a cold
mint plus an hour of diagnosis that had nothing to do with the code
under test.
Three defences:
· the phase-D revert is armed with a trap BEFORE the first mutation,
so it fires on any exit, Ctrl-C included;
· the revert is VERIFIED rather than fired into /dev/null, so a failed
unset can no longer masquerade as a successful one;
· the drill refuses to start on a host still carrying the mutations.
And drill/doctor.sh answers the question that kept being answered by
hand: what state is this host actually in? Network, profile, ACL,
leftover boxes, and whether a box can still resolve DNS — with --fix to
revert the leftovers.
RUNS.md gains trap 10.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 00:32:03 +00:00
dirty = ""
[ -n " $( incus profile device get claude-dev eth0 security.ipv4_filtering 2>/dev/null) " ] && dirty = " $dirty ipv4_filtering "
[ -n " $( incus profile device get claude-dev eth0 security.mac_filtering 2>/dev/null) " ] && dirty = " $dirty mac_filtering "
if [ -n " $dirty " ] ; then
2026-07-14 01:40:02 +00:00
note " this host carries the VETOED NIC filtering from an old rehearsal: $dirty — reverting "
fix(drill): stop poisoning the host, and add a doctor to prove it
Run 8's cold mint failed with 'cloud-init status: error' — the box could
not resolve deb.debian.org, or claude.ai, or anything. The cause was not
in that run at all: run 7's phase D set dns.mode=none on claudenet, the
run ended before reverting it, and every box minted afterwards came up
with no DNS.
This is the worst failure mode the drill has: a poisoned host does not
fail the next run honestly, it produces confident wrong answers. It is
how a false design veto against #16 got posted, and it wasted a cold
mint plus an hour of diagnosis that had nothing to do with the code
under test.
Three defences:
· the phase-D revert is armed with a trap BEFORE the first mutation,
so it fires on any exit, Ctrl-C included;
· the revert is VERIFIED rather than fired into /dev/null, so a failed
unset can no longer masquerade as a successful one;
· the drill refuses to start on a host still carrying the mutations.
And drill/doctor.sh answers the question that kept being answered by
hand: what state is this host actually in? Network, profile, ACL,
leftover boxes, and whether a box can still resolve DNS — with --fix to
revert the leftovers.
RUNS.md gains trap 10.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 00:32:03 +00:00
incus profile device unset claude-dev eth0 security.mac_filtering >/dev/null 2>& 1
incus profile device unset claude-dev eth0 security.ipv4_filtering >/dev/null 2>& 1
fi
fix(drill): clean the host BEFORE setup-host, and bound it
Run 6 stalled in setup-host.sh, which should be seconds on a host that
already has incus. The ordering was wrong: cleanup ran AFTER setup, so
setup-host reconfigured claudenet's ACLs while an aborted run's boxes
were still attached to that network — 'incus network set' then has to
push the change onto every live NIC. The same aborted run also left the
D-phase mutations (dns.mode=none, NIC filtering) in place, so setup was
converging against a moving target.
Boxes are now deleted and the mutations reverted first, so setup-host is
the no-op it should be. It is also bounded (5 min) and, on timeout,
prints the three things worth checking instead of hanging: instances
still on the network, the firewall unit, the incus daemon. Every cleanup
call gets its own timeout, so a wedged instance cannot stall the run
before the drill has printed a single line.
RUNS.md gains this as trap 8.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 23:09:51 +00:00
inf "clearing anything a previous run left behind…"
# One name at a time — 'incus delete -f a b c' aborts at the first MISSING name,
# which is how run 2 inherited run 1's boxes and cascaded five false FAILs.
2026-07-13 22:22:50 +00:00
for n in drill clone archive peer payroll cbprobe cbcopy cbnotours; do
fix(drill): clean the host BEFORE setup-host, and bound it
Run 6 stalled in setup-host.sh, which should be seconds on a host that
already has incus. The ordering was wrong: cleanup ran AFTER setup, so
setup-host reconfigured claudenet's ACLs while an aborted run's boxes
were still attached to that network — 'incus network set' then has to
push the change onto every live NIC. The same aborted run also left the
D-phase mutations (dns.mode=none, NIC filtering) in place, so setup was
converging against a moving target.
Boxes are now deleted and the mutations reverted first, so setup-host is
the no-op it should be. It is also bounded (5 min) and, on timeout,
prints the three things worth checking instead of hanging: instances
still on the network, the firewall unit, the incus daemon. Every cleanup
call gets its own timeout, so a wedged instance cannot stall the run
before the drill has printed a single line.
RUNS.md gains this as trap 8.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 23:09:51 +00:00
timeout -k 5 60 incus delete -f " $n " >/dev/null 2>& 1
2026-07-13 22:22:50 +00:00
done
fix(drill): clean the host BEFORE setup-host, and bound it
Run 6 stalled in setup-host.sh, which should be seconds on a host that
already has incus. The ordering was wrong: cleanup ran AFTER setup, so
setup-host reconfigured claudenet's ACLs while an aborted run's boxes
were still attached to that network — 'incus network set' then has to
push the change onto every live NIC. The same aborted run also left the
D-phase mutations (dns.mode=none, NIC filtering) in place, so setup was
converging against a moving target.
Boxes are now deleted and the mutations reverted first, so setup-host is
the no-op it should be. It is also bounded (5 min) and, on timeout,
prints the three things worth checking instead of hanging: instances
still on the network, the firewall unit, the incus daemon. Every cleanup
call gets its own timeout, so a wedged instance cannot stall the run
before the drill has printed a single line.
RUNS.md gains this as trap 8.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 23:09:51 +00:00
if incus network show claudenet >/dev/null 2>& 1; then
timeout -k 5 30 incus network unset claudenet dns.mode >/dev/null 2>& 1
fi
if incus profile show claude-dev >/dev/null 2>& 1; then
timeout -k 5 30 incus profile device unset claude-dev eth0 security.mac_filtering >/dev/null 2>& 1
timeout -k 5 30 incus profile device unset claude-dev eth0 security.ipv4_filtering >/dev/null 2>& 1
fi
left = " $( incus list --format csv --columns n 2>/dev/null | tr '\n' ' ' ) "
[ -n " $left " ] && inf " instances still on this host (not ours, left alone): $left "
inf "running setup-host.sh (in-group pass: network, ACL, profile, firewall)…"
if ! timeout -k 10 300 ~/.local/share/claudebox/host/setup-host.sh; then
echo "drill: setup-host.sh failed or timed out (>5 min)." >& 2
echo " it should take seconds on a host that already has incus. usual causes:" >& 2
echo " · instances still attached to claudenet while its ACLs are reconfigured" >& 2
echo " incus list" >& 2
echo " · the firewall unit not completing" >& 2
echo " systemctl status claudebox-firewall.service --no-pager" >& 2
echo " · the incus daemon wedged by an earlier aborted run" >& 2
echo " systemctl status incus --no-pager; journalctl -u incus -n 30 --no-pager" >& 2
exit 1
fi
inf "host setup complete"
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
# A real server has room for the production profile (8GiB/4cpu), and drilling the
# real profile is worth more than drilling a shrunken one. Only shrink if we must.
ram = " $( awk '/MemTotal/{print int($2/1024/1024)}' /proc/meminfo) "
if [ " $ram " -lt 20 ] ; then
incus profile set claude-dev limits.memory= 3GiB limits.cpu= 2
note " host has ${ ram } GiB RAM — lowered claude-dev to 3GiB/2cpu for the drill (production profile is 8GiB/4cpu, and that is what was NOT drilled) "
else
inf " host has ${ ram } GiB RAM — drilling the production profile (8GiB/4cpu) unchanged "
fi
KVM = 0; [ -e /dev/kvm ] && KVM = 1
[ " $KVM " = 1 ] && inf "/dev/kvm present — boxes will be VMs (the real trust boundary)" \
|| note "NO /dev/kvm on this host — claudebox will fall back to CONTAINER mode, so this run does NOT validate the VM trust boundary"
# ===========================================================================
phase "A. Incus semantics — the assumptions claudebox is built on"
# ===========================================================================
incus launch images:debian/13 cbprobe --config user.claudebox= 1 >/dev/null 2>& 1
incus launch images:debian/13 cbnotours >/dev/null 2>& 1 # untagged: not ours
sleep 3
# A1 — the tag read. #13 puts this on the path of EVERY box command.
t = " $( incus config get cbprobe user.claudebox 2>& 1) "
[ " $t " = "1" ] && ok "config get user.claudebox → '1'" \
|| no " config get user.claudebox → ' $t ' (expected '1'; every box command would fail closed) "
# A2 — the list filter, and that it EXCLUDES an instance we didn't mint
f = " $( incus list user.claudebox= 1 --format csv --columns nstS 2>& 1) "
if echo " $f " | grep -q '^cbprobe,' && ! echo " $f " | grep -q '^cbnotours,' ; then
ok "list filter user.claudebox=1 selects ours, excludes theirs"
else
no " list filter user.claudebox=1 is wrong — got: $( echo " $f " | tr '\n' ' ' ) "
fi
# A3 — four fields, no commas/newlines to mangle the awk table
n = " $( echo " $f " | grep '^cbprobe,' | awk -F, '{print NF}' ) "
[ " $n " = 4 ] && ok "--columns nstS → 4 clean CSV fields" || no " --columns nstS → $n fields (the list table would garble) "
# A4 — the state string require_stopped compares against
s = " $( incus list cbprobe --format csv --columns s 2>& 1 | head -1) "
[ " $s " = RUNNING ] && ok "state column → 'RUNNING'" || no " state column → ' $s ' (require_stopped compares against RUNNING/STOPPED) "
# A5 — does rename REFUSE a running instance? #13's precondition bets it does.
if r = " $( incus rename cbprobe cbprobe2 2>& 1) " ; then
no "incus renamed a RUNNING instance — #13's 'stopped' precondition is unnecessary (merely conservative)"
incus rename cbprobe2 cbprobe >/dev/null 2>& 1
else
ok " incus refuses to rename a running instance → $( echo " $r " | head -1 | cut -c1-60) "
fi
# A6 — snapshot list CSV: 'info' reads field 1 as the label
incus snapshot create cbprobe authed >/dev/null 2>& 1
s1 = " $( incus snapshot list cbprobe --format csv 2>& 1 | head -1) "
[ " $( echo " $s1 " | cut -d, -f1) " = authed ] && ok "snapshot list csv → field 1 is the label" \
|| no " snapshot list csv field 1 ≠ label — got: $s1 "
# A7 — the IPv4 column. #9 assumes it can be quoted/multi-line, hence fetching it apart.
inf " ipv4 column raw: $( incus list cbprobe --format csv --columns 4 2>& 1 | tr '\n' '|' ) "
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
# A8 — unset config keys read as EMPTY with exit 0 (#15 B4). The '|| echo root'
# fallback #12 first proposed could never fire if so; #17's lookup depends on this.
u = " $( incus config get cbprobe user.never-set 2>& 1) " ; rc = $?
if [ " $rc " -eq 0 ] && [ -z " $u " ] ; then
ok "config get on an unset key → empty string, exit 0"
aud "B4 config-get unset key: empty + exit 0 — #17 fallbacks must use \${var:-}, never ||"
else
note " config get on an unset key → rc= $rc out=' $u ' (not the documented empty+0 — #17's lookup adapts) "
aud " B4 config-get unset key: rc= $rc out=' $u ' "
fi
# A9 — 'incus copy' preserves user.* keys (#15 B2). #17's whole metadata design:
# a clone must still know what it is without consulting the template.
incus config set cbprobe user.box.user claude 2>/dev/null
incus stop -f cbprobe >/dev/null 2>& 1
incus copy cbprobe cbcopy >/dev/null 2>& 1
c = " $( incus config get cbcopy user.box.user 2>/dev/null) "
if [ " $c " = claude ] ; then
ok "incus copy preserves user.* keys (a clone knows what it is)"
aud "B2 copy preserves user.*: YES — #17's metadata-stamp design holds"
else
no " incus copy DROPPED user.* keys (got ' $c ') — #17's metadata design fails without them "
aud "B2 copy preserves user.*: NO — #17 blocked as designed"
fi
incus delete -f cbcopy >/dev/null 2>& 1
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
incus delete -f cbprobe cbnotours >/dev/null 2>& 1
# ===========================================================================
phase "B. The claudebox surface"
# ===========================================================================
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
# Compare against the installed tree's VERSION file, not a hardcoded number —
# a pinned literal here would fail the drill on every release.
expected = " $( cat " $HOME /.local/share/claudebox/VERSION " 2>/dev/null || echo '?' ) "
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
v = " $( claudebox --version 2>& 1) "
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
case " $v " in *" $expected " *) ok " claudebox --version → $v " ; ; *) no " version mismatch: CLI says ' $v ', VERSION file says ' $expected ' " ; ; esac
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
claudebox list >/dev/null 2>& 1 && claudebox list 2>& 1 | grep -q 'no boxes yet' \
&& ok "empty host: 'no boxes yet', exit 0" || no "empty-host message wrong"
printf '\n minting a box (cold, ~10 min)…\n'
t0 = $SECONDS
if claudebox new --name drill >/tmp/new.log 2>& 1; then
ok " claudebox new --name drill ( $(( SECONDS - t0)) s) "
else
no " claudebox new FAILED — tail: $( tail -3 /tmp/new.log | tr '\n' ' ' ) "
echo; echo "── cannot continue without a box" ; printf ' %s\n' " ${ findings [@] } " ; exit 1
fi
typ = " $( claudebox list | awk '$1 == "drill" { print $3 }' ) "
if [ " $KVM " = 1 ] ; then
[ " $typ " = VM ] && ok "the box is a VM — the trust boundary is real" \
|| no " the box is ' $typ ' but /dev/kvm exists — it should have been a VM "
else
note " the box is ' $typ ' (no /dev/kvm on this host) "
fi
claudebox info drill | grep -q '^IPV4' && ok "info shows an IPv4" || no "info has no IPV4 row"
claudebox info drill | grep -q 'SNAPSHOTS (none)' && ok "info: no snapshots yet, offers to take one" || no "info snapshot-empty state wrong"
2026-07-13 22:05:32 +00:00
if claudebox exec drill -- claude --version >/dev/null 2>& 1; then
ok "Claude Code is installed in the box"
elif timeout 30 claudebox exec drill -- bash -lc 'claude --version' >/dev/null 2>& 1; then
no "'claude' is installed but NOT on exec's PATH — repo bug: the help promises 'claudebox exec work -- claude --version'"
inf " PATH as exec sees it: $( timeout 30 claudebox exec drill -- printenv PATH 2>/dev/null) "
else
no "'claude --version' failed inside the box"
2026-07-13 22:22:50 +00:00
# diag output must skip the hatch's own 'claudebox: incus exec …' announce lines
hatch_out( ) { timeout 30 claudebox incus drill -- exec { } -- " $@ " 2>& 1 | grep -v '^claudebox:' | tail -1 | cut -c1-120; }
inf " cloud-init: $( hatch_out cloud-init status) "
inf " binary runs? $( hatch_out sudo -u claude /home/claude/.local/bin/claude --version) "
inf " exec PATH: $( timeout 30 claudebox exec drill -- printenv PATH 2>/dev/null | tail -1) "
2026-07-13 22:05:32 +00:00
fi
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
claudebox exec drill -- gh --version >/dev/null 2>& 1 \
&& ok "the GitHub CLI is installed in the box (PR #5)" || no "'gh --version' failed inside the box"
# --- the snapshot → clone workflow, which is the whole point of the tool ---
claudebox snapshot drill authed 2>& 1 | grep -q authed && ok "snapshot drill authed" || no "snapshot failed"
claudebox info drill | grep -q 'authed' && ok "info lists the snapshot label" || no "info does not show the label"
claudebox info drill | grep -q -- '--from drill/authed' && ok "info prints the --from line to clone it" || no "info lacks the --from hint"
# --- the boundary: an instance claudebox did NOT mint ----------------------
incus launch images:debian/13 payroll >/dev/null 2>& 1 # somebody else's instance
sleep 2
claudebox down payroll 2>& 1 | grep -q 'no such box' && ok "boundary: 'down' refuses an untagged instance" || no "boundary: 'down' touched an instance claudebox didn't mint!"
claudebox rm payroll --force 2>& 1 | grep -q 'no such box' && ok "boundary: 'rm' refuses an untagged instance" || no "boundary: 'rm' would DELETE a foreign instance!"
claudebox incus payroll -- config show 2>& 1 | grep -q 'no such box' && ok "boundary: the escape hatch refuses it too" || no "boundary: the hatch reached a foreign instance!"
incus list payroll --format csv --columns ns | grep -q '^payroll,RUNNING' && ok "…and payroll is still running, untouched" || no "payroll was harmed — the boundary leaked"
incus delete -f payroll >/dev/null 2>& 1
# --- rename, and its precondition -----------------------------------------
claudebox rename drill archive 2>& 1 | grep -qi 'RUNNING' && ok "rename refuses a running box, and says how to fix it" || no "rename did not refuse a running box"
claudebox down drill >/dev/null 2>& 1 && ok "down drill" || no "down failed"
claudebox rename drill archive 2>& 1 | grep -q 'renamed drill to archive' && ok "rename drill → archive (stopped)" || no "rename failed on a stopped box"
claudebox list | grep -q '^archive' && ok "list shows the new name" || no "list still shows the old name"
claudebox info archive | grep -q authed && ok "the snapshot followed the rename" || no "snapshot lost across the rename"
# --- clone from a snapshot of a renamed box --------------------------------
printf '\n cloning from the snapshot…\n'
if claudebox new --name clone --from archive/authed >/tmp/clone.log 2>& 1; then
ok "new --from archive/authed (clone of a snapshot of a renamed box)"
claudebox exec clone -- true >/dev/null 2>& 1 && ok "the clone is alive and enterable" || no "the clone is not enterable"
else
no " clone FAILED — tail: $( tail -3 /tmp/clone.log | tr '\n' ' ' ) "
fi
# --- the escape hatch ------------------------------------------------------
claudebox incus archive -- config show 2>/dev/null | grep -q 'user.claudebox' && ok "hatch: 'incus archive -- config show', instance appended" || no "hatch passthrough failed"
h = " $( claudebox incus archive -- config device add { } scratch disk source = /tmp path = /mnt/scratch 2>& 1) "
echo " $h " | grep -q 'isolation stack' && ok "hatch warns when a command can break isolation" || no "hatch did not warn on a device add"
claudebox incus archive -- config device remove { } scratch >/dev/null 2>& 1
# --- rm, and the guard that did not used to exist --------------------------
claudebox rm clone </dev/null 2>& 1 | grep -q 'refusing' && ok "rm with no TTY and no --force refuses (exit 2)" || no "rm destroyed a box with no confirmation!"
claudebox rm clone --force 2>& 1 | grep -q 'removed' && ok "rm --force removes the clone" || no "rm --force failed"
# --- the CLI contract ------------------------------------------------------
claudebox lst 2>& 1 | grep -q "did you mean 'list'" && ok "typo → did-you-mean, exit 2" || no "unknown command not suggested"
claudebox list archive 2>& 1 | grep -q 'claudebox info archive' && ok "'list <box>' points at info" || no "'list <box>' does not point at info"
claudebox snapshot archive --labl x 2>& 1 | grep -q 'unknown option' && ok "typo'd flag rejected (not swallowed as a label)" || no "unknown flag was swallowed"
# ===========================================================================
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
phase "C. Isolation baseline — does the boundary actually hold? (#15 section A)"
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
# ===========================================================================
claudebox start archive >/dev/null 2>& 1
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
wait_box archive && ok "archive is back up (agent answering)" \
|| no "archive did not come back within 2 min of start"
# Sibling isolation needs a sibling. Clone from the snapshot — fast, no cold mint.
printf '\n cloning a peer for the sibling probes…\n'
if claudebox new --name peer --from archive/authed >/tmp/peer.log 2>& 1 && wait_box peer; then
ok "peer minted from archive/authed and answering"
else
no " peer clone failed or never answered — tail: $( tail -3 /tmp/peer.log | tr '\n' ' ' ) "
fi
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
# C1 — public egress (#15 A1; resolving the hostname also proves A5, gateway DNS)
2026-07-14 00:04:01 +00:00
BASELINE_OK = 1
fix(drill): read curl's message — the exit code cannot tell you what happened
Boxes are isolated. The kernel says so: 'isolated on' on both live bridge
ports. A box gets 100% packet loss pinging its sibling, and its TCP
connect burns the full timeout and dies with "Could not connect to
server" — not "Connection refused", which is what an arriving packet
gets, instantly.
The drill called that a FAIL. curl exit 7 is "failed to connect", and it
covers BOTH a refusal (a RST came back — the packet ARRIVED) and an
unreachable host (nothing came back — it was DROPPED). Opposite
conclusions, one exit code, and the drill mapped 7 → "arrived". So for
two runs after the isolation fix had landed and was working, the drill
reported a working boundary as a broken one.
The words distinguish what the number cannot. box_probe now returns
reachable | refused | dropped by reading the message, ping corroborates,
and every isolation probe (sibling, box→host, RFC1918, host→box) uses it.
RUNS.md gains trap 12. It is the same disease as all eleven before it:
trusting a proxy for the fact instead of the fact.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 02:33:02 +00:00
if [ " $( box_probe archive https://api.github.com 20) " = reachable ] ; then
2026-07-14 00:04:01 +00:00
ok "box reaches the public internet (and gateway DNS resolves public names)"
aud "A1/A5 egress + public DNS: PASS"
else
BASELINE_OK = 0
no "box cannot reach the internet (a box that can't is useless)"
aud "A1/A5 egress: FAIL"
fi
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
2026-07-13 23:02:48 +00:00
# C2 — box → host (#15 A2). The host DOES listen on the gateway: dnsmasq is on
# :53 by design (that carve-out is what makes egress DNS work). So probe a port
# nothing serves and read refused-vs-dropped — refused would mean the box's
# packet reached the host's stack, which is the thing the firewall must prevent.
# (No background listener: one less process to leak, one less way to wedge.)
fix(drill): read curl's message — the exit code cannot tell you what happened
Boxes are isolated. The kernel says so: 'isolated on' on both live bridge
ports. A box gets 100% packet loss pinging its sibling, and its TCP
connect burns the full timeout and dies with "Could not connect to
server" — not "Connection refused", which is what an arriving packet
gets, instantly.
The drill called that a FAIL. curl exit 7 is "failed to connect", and it
covers BOTH a refusal (a RST came back — the packet ARRIVED) and an
unreachable host (nothing came back — it was DROPPED). Opposite
conclusions, one exit code, and the drill mapped 7 → "arrived". So for
two runs after the isolation fix had landed and was working, the drill
reported a working boundary as a broken one.
The words distinguish what the number cannot. box_probe now returns
reachable | refused | dropped by reading the message, ping corroborates,
and every isolation probe (sibling, box→host, RFC1918, host→box) uses it.
RUNS.md gains trap 12. It is the same disease as all eleven before it:
trusting a proxy for the fact instead of the fact.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 02:33:02 +00:00
hv = " $( box_probe archive http://10.87.0.1:8099) "
2026-07-13 23:02:48 +00:00
case " $hv " in
reachable| refused)
no " THE BOX'S PACKETS REACH THE HOST on 10.87.0.1:8099 [ $hv ] — the firewall rules are not holding "
aud " A2 box→host: FAIL — $hv (the packet reached the host's stack) " ; ;
dropped)
ok "box → host is blocked (no path to the machine's sockets)"
aud "A2 box→host: dropped" ; ;
*)
note " box→host probe inconclusive ( $hv ) "
aud " A2 box→host: INCONCLUSIVE ( $hv ) " ; ;
esac
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
# C3 — RFC1918 (#15 A2)
fix(drill): read curl's message — the exit code cannot tell you what happened
Boxes are isolated. The kernel says so: 'isolated on' on both live bridge
ports. A box gets 100% packet loss pinging its sibling, and its TCP
connect burns the full timeout and dies with "Could not connect to
server" — not "Connection refused", which is what an arriving packet
gets, instantly.
The drill called that a FAIL. curl exit 7 is "failed to connect", and it
covers BOTH a refusal (a RST came back — the packet ARRIVED) and an
unreachable host (nothing came back — it was DROPPED). Opposite
conclusions, one exit code, and the drill mapped 7 → "arrived". So for
two runs after the isolation fix had landed and was working, the drill
reported a working boundary as a broken one.
The words distinguish what the number cannot. box_probe now returns
reachable | refused | dropped by reading the message, ping corroborates,
and every isolation probe (sibling, box→host, RFC1918, host→box) uses it.
RUNS.md gains trap 12. It is the same disease as all eleven before it:
trusting a proxy for the fact instead of the fact.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 02:33:02 +00:00
case " $( box_probe archive http://192.168.1.1) " in
reachable| refused)
no "box REACHED a private-range address — the ACL is not dropping RFC1918"
aud "A2 RFC1918: FAIL" ; ;
*)
ok "box → RFC1918 is dropped by the ACL"
aud "A2 RFC1918: dropped" ; ;
esac
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
2026-07-13 23:02:48 +00:00
# C4 — SIBLING isolation (#15 A3): the central claim of #12, and the one probe
# three runs failed to fire. NO listener on the peer, deliberately — a closed
# port answers the question just as well (refused = the packet arrived), and
# the listener was what kept wedging the run. Ping corroborates: if the two
# disagree, say so rather than pick one.
fix(drill): the NIC inside a VM is enp5s0, not eth0 — read by subnet instead
A3, the one probe the whole audit exists for, has never fired in six
runs. It was never the network: the profile names the DEVICE eth0, but
inside a VM guest predictable naming renames it enp5s0, so every address
lookup — first the '(eth0)' CSV match, then 'ip addr show dev eth0' —
was hunting an interface that does not exist. I fixed that symptom twice
without ever questioning the assumption underneath it.
Read the address from inside the box and select by SUBNET (10.87.x, what
claudenet hands out) rather than by interface name. docker0's 172.17.x
is the decoy; the NIC's name is the guest's business, not ours.
A3 also gains a guard it should have had from the start: if the peer and
the source hold the SAME address, refuse to probe. That is not
hypothetical — clones were inheriting their source's machine-id, hence
its DHCP lease, hence its address, so 'archive → peer' was archive
probing itself and would have reported a cheerful 'reachable' as an
isolation failure.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 23:32:22 +00:00
PEER_IP = " $( claudenet_ip peer) "
ARCH_IP_PRE = " $( claudenet_ip archive) "
if [ -n " $PEER_IP " ] && [ " $PEER_IP " = " $ARCH_IP_PRE " ] ; then
# Guard, because this actually happened: a clone inherited its source's
# machine-id, hence its DHCP lease, hence its ADDRESS. Probing "archive →
# peer" was archive probing itself, and would have reported a cheerful
# "reachable" as a sibling-isolation failure. Never let A3 answer this.
no " archive and peer hold the SAME address ( $PEER_IP ) — the clone did not get its own identity; A3 cannot be probed "
aud "A3 sibling: NOT PROBED — clone/source IP collision (see the clone-identity fix)"
elif [ -n " $PEER_IP " ] ; then
fix(drill): read curl's message — the exit code cannot tell you what happened
Boxes are isolated. The kernel says so: 'isolated on' on both live bridge
ports. A box gets 100% packet loss pinging its sibling, and its TCP
connect burns the full timeout and dies with "Could not connect to
server" — not "Connection refused", which is what an arriving packet
gets, instantly.
The drill called that a FAIL. curl exit 7 is "failed to connect", and it
covers BOTH a refusal (a RST came back — the packet ARRIVED) and an
unreachable host (nothing came back — it was DROPPED). Opposite
conclusions, one exit code, and the drill mapped 7 → "arrived". So for
two runs after the isolation fix had landed and was working, the drill
reported a working boundary as a broken one.
The words distinguish what the number cannot. box_probe now returns
reachable | refused | dropped by reading the message, ping corroborates,
and every isolation probe (sibling, box→host, RFC1918, host→box) uses it.
RUNS.md gains trap 12. It is the same disease as all eleven before it:
trusting a proxy for the fact instead of the fact.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 02:33:02 +00:00
inf " probing archive ( $ARCH_IP_PRE ) → peer ( $PEER_IP ): a REFUSAL means it arrived; silence means it was dropped "
v = " $( box_probe archive " http:// $PEER_IP :8088 " ) "
box_pings archive " $PEER_IP " ; png = $?
if [ " $v " = reachable ] || [ " $v " = refused ] ; then
no " BOX A REACHES BOX B ( $PEER_IP ) — sibling isolation does NOT hold [tcp: $v ] "
aud " A3 sibling: FAIL — tcp $v (the packet arrived) "
elif [ " $png " -eq 0 ] ; then
no "TCP to box B goes nowhere, but it ANSWERS ICMP — sibling isolation is only partial"
aud "A3 sibling: PARTIAL — tcp dropped, ping replies"
else
ok "box A cannot reach box B: TCP goes nowhere, ICMP unanswered"
aud "A3 sibling: BLOCKED — tcp dropped + no icmp reply (security.port_isolation)"
fi
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
else
fix(drill): the NIC inside a VM is enp5s0, not eth0 — read by subnet instead
A3, the one probe the whole audit exists for, has never fired in six
runs. It was never the network: the profile names the DEVICE eth0, but
inside a VM guest predictable naming renames it enp5s0, so every address
lookup — first the '(eth0)' CSV match, then 'ip addr show dev eth0' —
was hunting an interface that does not exist. I fixed that symptom twice
without ever questioning the assumption underneath it.
Read the address from inside the box and select by SUBNET (10.87.x, what
claudenet hands out) rather than by interface name. docker0's 172.17.x
is the decoy; the NIC's name is the guest's business, not ours.
A3 also gains a guard it should have had from the start: if the peer and
the source hold the SAME address, refuse to probe. That is not
hypothetical — clones were inheriting their source's machine-id, hence
its DHCP lease, hence its address, so 'archive → peer' was archive
probing itself and would have reported a cheerful 'reachable' as an
isolation failure.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 23:32:22 +00:00
no "could not read peer's claudenet address — the sibling probe never ran"
aud "A3 sibling: NOT PROBED (no 10.87.x address on peer)"
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
fi
2026-07-14 01:29:33 +00:00
# C5 — DNS enumeration (#15 A4). Now a CONTRACT, not an observation: setup-host
# sets dns.mode=none precisely so a box cannot enumerate its siblings.
fix(drill): no in-box probe can hang the run again
Run 4 hung 10+ minutes at C4. The cause is a shell trap, not a network
one: $( ) waits for stdout to CLOSE, not for the command to exit, so a
grandchild inheriting the exec session's stdout holds the substitution
open forever — and 'timeout' does not save you, because it kills the
wrapper, not the process holding the pipe. 'claudebox exec' made it
likelier still: it becomes 'sudo -u claude -i', a login zsh with
oh-my-zsh, which is a lot of machinery to keep a fd alive.
Every in-box read now goes through in_box()/box_curl(): incus exec
directly, stdin pinned to /dev/null, output landed in a file rather than
a pipe, and 'timeout -k' to hard-kill. box_curl returns the curl exit
code as its finding — for A3 the exit code IS the answer (0 connected,
7 refused-so-it-arrived, 28 dropped). C4 also prints what it is probing
before it probes, so a stall is locatable instead of silent.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 22:55:43 +00:00
e1 = " $( in_box archive getent hosts peer) "
e2 = " $( in_box archive getent hosts peer.incus) "
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
if [ -n " $e1 $e2 " ] ; then
2026-07-14 01:29:33 +00:00
no " a box can still RESOLVE its sibling ( $( printf '%s' " $e1 $e2 " | head -1 | cut -c1-40) ) — dns.mode=none is not holding "
aud "A4 dns enumeration: LEAKS — the fix is not in effect"
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
else
2026-07-14 01:29:33 +00:00
ok "a box cannot resolve its sibling's name (no DNS enumeration)"
aud "A4 dns enumeration: blocked (dns.mode=none)"
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
fi
# C6 — IPv6 off (#15 A6): every ACL rule is IPv4-only; off is the only cover.
[ " $( incus network get claudenet ipv6.address 2>/dev/null) " = none ] \
&& { ok "claudenet ipv6.address = none (the IPv4-only ACLs have no uncovered path)" ; aud "A6 ipv6: none, as contract requires" ; } \
|| { no "claudenet has IPv6 enabled — and not one ACL rule covers IPv6" ; aud "A6 ipv6: ENABLED and uncovered" ; }
2026-07-13 23:02:48 +00:00
# C7 — inbound, host → box (#15 A7): the ACL's default ingress drop. Same
# listener-free logic, run from the host this time.
fix(drill): the NIC inside a VM is enp5s0, not eth0 — read by subnet instead
A3, the one probe the whole audit exists for, has never fired in six
runs. It was never the network: the profile names the DEVICE eth0, but
inside a VM guest predictable naming renames it enp5s0, so every address
lookup — first the '(eth0)' CSV match, then 'ip addr show dev eth0' —
was hunting an interface that does not exist. I fixed that symptom twice
without ever questioning the assumption underneath it.
Read the address from inside the box and select by SUBNET (10.87.x, what
claudenet hands out) rather than by interface name. docker0's 172.17.x
is the decoy; the NIC's name is the guest's business, not ours.
A3 also gains a guard it should have had from the start: if the peer and
the source hold the SAME address, refuse to probe. That is not
hypothetical — clones were inheriting their source's machine-id, hence
its DHCP lease, hence its address, so 'archive → peer' was archive
probing itself and would have reported a cheerful 'reachable' as an
isolation failure.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 23:32:22 +00:00
ARCH_IP = " $( claudenet_ip archive) "
2026-07-13 23:02:48 +00:00
if [ -n " $ARCH_IP " ] ; then
fix(drill): read curl's message — the exit code cannot tell you what happened
Boxes are isolated. The kernel says so: 'isolated on' on both live bridge
ports. A box gets 100% packet loss pinging its sibling, and its TCP
connect burns the full timeout and dies with "Could not connect to
server" — not "Connection refused", which is what an arriving packet
gets, instantly.
The drill called that a FAIL. curl exit 7 is "failed to connect", and it
covers BOTH a refusal (a RST came back — the packet ARRIVED) and an
unreachable host (nothing came back — it was DROPPED). Opposite
conclusions, one exit code, and the drill mapped 7 → "arrived". So for
two runs after the isolation fix had landed and was working, the drill
reported a working boundary as a broken one.
The words distinguish what the number cannot. box_probe now returns
reachable | refused | dropped by reading the message, ping corroborates,
and every isolation probe (sibling, box→host, RFC1918, host→box) uses it.
RUNS.md gains trap 12. It is the same disease as all eleven before it:
trusting a proxy for the fact instead of the fact.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 02:33:02 +00:00
hmsg = " $( curl -sS -m 5 -o /dev/null " http:// $ARCH_IP :8087 " 2>& 1) " ; hrc = $?
if [ " $hrc " -eq 0 ] ; then hv = reachable
elif printf '%s' " $hmsg " | grep -q 'Connection refused' ; then hv = refused
else hv = dropped
fi
2026-07-13 23:02:48 +00:00
case " $hv " in
reachable| refused)
no " the HOST's packets REACH the box ( $ARCH_IP ) — the default ingress drop is not holding [ $hv ] "
aud " A7 inbound host→box: FAIL — $hv (the packet arrived) " ; ;
dropped)
ok "host → box is dropped (entry is 'incus exec' only, as designed)"
aud "A7 inbound host→box: dropped" ; ;
*)
note " inbound probe inconclusive ( $hv ) "
aud " A7 inbound host→box: INCONCLUSIVE ( $hv ) " ; ;
esac
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
else
fix(drill): the NIC inside a VM is enp5s0, not eth0 — read by subnet instead
A3, the one probe the whole audit exists for, has never fired in six
runs. It was never the network: the profile names the DEVICE eth0, but
inside a VM guest predictable naming renames it enp5s0, so every address
lookup — first the '(eth0)' CSV match, then 'ip addr show dev eth0' —
was hunting an interface that does not exist. I fixed that symptom twice
without ever questioning the assumption underneath it.
Read the address from inside the box and select by SUBNET (10.87.x, what
claudenet hands out) rather than by interface name. docker0's 172.17.x
is the decoy; the NIC's name is the guest's business, not ours.
A3 also gains a guard it should have had from the start: if the peer and
the source hold the SAME address, refuse to probe. That is not
hypothetical — clones were inheriting their source's machine-id, hence
its DHCP lease, hence its address, so 'archive → peer' was archive
probing itself and would have reported a cheerful 'reachable' as an
isolation failure.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 23:32:22 +00:00
no "could not read archive's claudenet address — the inbound probe never ran"
2026-07-13 23:02:48 +00:00
aud "A7 inbound host→box: NOT PROBED"
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
fi
# ===========================================================================
2026-07-14 01:29:33 +00:00
phase "D. The isolation contract, stated"
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
# ===========================================================================
2026-07-14 01:29:33 +00:00
# Phase D used to REHEARSE the hardening on a throwaway host, because nobody
# knew whether it would work. That question is settled: the hardening now ships
# in setup-host.sh and claudebox-firewall.sh, so phase C tests the real thing
# and there is nothing left to rehearse. What the rehearsal established, kept
# here so it is not re-litigated:
#
# · @internal is REJECTED as an ACL destination on a bridge network
# ("Unsupported nftables subject") — so the sibling drop is an nftables
# bridge-family rule, not an ACL rule. It has to be: an L3 ACL never sees
# frames switched between two ports of one bridge, which is why box→box was
# wide open while the ACL looked airtight.
# · dns.mode=none closes the enumeration leak and public egress survives it.
# · security.ipv4_filtering BREAKS the box's networking (dockerd comes up but
# cannot pull or run a container). VETOED — it is not in the shipped stack.
2026-07-14 00:04:01 +00:00
#
2026-07-14 01:29:33 +00:00
inf "@internal: unsupported on bridge ACLs ⇒ the sibling drop is an nft bridge rule"
inf "dns.mode=none: shipped (closes DNS enumeration, egress unaffected)"
inf "security.ipv4_filtering: VETOED — it breaks the box. Not shipped."
inf "the contract is now tested in phase C against the real stack, not rehearsed"
# The lesson that cost the most: a verdict measured on a broken box is not a
# verdict. Run 7 reported "L2 filtering BREAKS the box" from a box whose network
# was already dead, and #16 was nearly redesigned around it. If the baseline
# failed, say plainly that phase C's isolation results cannot be trusted.
2026-07-14 00:04:01 +00:00
if [ " $BASELINE_OK " -ne 1 ] ; then
2026-07-14 01:29:33 +00:00
no "the box could not reach the internet AT ALL — every isolation result above is suspect, not a pass"
inf "a boundary that 'holds' on a box with no network holds nothing. fix the baseline, re-run."
inf "start with: bash drill/doctor.sh"
fix(drill): stop poisoning the host, and add a doctor to prove it
Run 8's cold mint failed with 'cloud-init status: error' — the box could
not resolve deb.debian.org, or claude.ai, or anything. The cause was not
in that run at all: run 7's phase D set dns.mode=none on claudenet, the
run ended before reverting it, and every box minted afterwards came up
with no DNS.
This is the worst failure mode the drill has: a poisoned host does not
fail the next run honestly, it produces confident wrong answers. It is
how a false design veto against #16 got posted, and it wasted a cold
mint plus an hour of diagnosis that had nothing to do with the code
under test.
Three defences:
· the phase-D revert is armed with a trap BEFORE the first mutation,
so it fires on any exit, Ctrl-C included;
· the revert is VERIFIED rather than fired into /dev/null, so a failed
unset can no longer masquerade as a successful one;
· the drill refuses to start on a host still carrying the mutations.
And drill/doctor.sh answers the question that kept being answered by
hand: what state is this host actually in? Network, profile, ACL,
leftover boxes, and whether a box can still resolve DNS — with --fix to
revert the leftovers.
RUNS.md gains trap 10.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 00:32:03 +00:00
fi
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
# ===========================================================================
if [ " $KEEP " = 1 ] ; then
phase "Boxes left up (--keep-boxes)"
claudebox list
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
inf "note: the D-phase mutations (dns.mode=none, NIC filtering) are still applied"
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
else
2026-07-13 22:22:50 +00:00
# every name the drill can have left, whatever branch a partial run took
for n in drill clone archive peer; do claudebox rm " $n " --force >/dev/null 2>& 1; done
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
claudebox list 2>& 1 | grep -q 'no boxes yet' && ok "teardown: no boxes left" || no "a box survived teardown"
fi
phase "Summary"
printf ' %s passed, %s failed\n' " $pass " " $fail "
if [ " ${# findings [@] } " -gt 0 ] ; then
echo
printf ' %s\n' " ${ findings [@] } "
fi
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
if [ " ${# audit [@] } " -gt 0 ] ; then
phase "#15 audit answers — paste this block into heavy-duty/claudebox#15"
printf ' %s\n' " ${ audit [@] } "
fi
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
echo
test: fold the #15 audit into the drill — sibling probes and a hardening rehearsal
The drill covered the CLI surface and basic egress but never probed the
one claim everything hinges on: it only ever had one box running, so
sibling isolation — #12's central, never-reproduced finding — went
untested. Phase C now clones a peer and probes box-to-box reachability
(with a listener, so "refused" cannot masquerade as "dropped"), DNS
enumeration, IPv6-off, and host-to-box inbound. A new phase D applies
the exact changes #16 proposes (dns.mode=none, NIC filtering, @internal)
and watches what breaks — a FAIL there is a design veto caught before
the code is written. The run ends with an answers block to paste into
issue #15.
Also: defaults now install heavy-duty/claudebox@main (the old default
pointed at a fork branch that stopped moving when #13 merged), the
version check compares against the installed VERSION file instead of a
hardcoded 0.3.0, and the blind sleep after start became an agent poll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 21:30:47 +00:00
inf "this host still has Incus, claudenet, the ACL, the profile and the firewall rules"
inf "(plus, unless re-run: dns.mode=none and NIC filtering from the D phase)."
test: add an end-to-end drill against a real Incus
The repo has no tests and no CI, and everything in #9/#10/#13 was verified
against a stubbed incus. The interesting failures aren't in the bash — they're
in what Incus actually does, which is exactly what a stub gets wrong. So the
drill runs the real thing on a host you can format.
A. Incus semantics — the assumptions claudebox is built on, probed directly:
the user.claudebox=1 tag read (now on the path of every box command in
#13), the list filter, the CSV shape, the state string, whether rename
really refuses a running instance, the snapshot-list field order.
B. The claudebox surface — mint, list, info, snapshot, clone-from-snapshot-of-
a-renamed-box, rename, the escape hatch, the rm guard, the CLI contract —
and the boundary: an instance claudebox did NOT mint has down/rm/hatch
aimed at it, must refuse all three, and must still be standing after.
C. Isolation — from inside a real box: public egress works, box → host is
blocked, box → RFC1918 is dropped.
Destructive by design, and it says so before it starts: it asks for consent
(--yes to skip), and refuses outright with no TTY. Without /dev/kvm it reports
that the VM trust boundary was NOT validated rather than passing on a container.
2026-07-13 21:15:56 +00:00
inf "to undo: ~/.local/share/claudebox/host/teardown-host.sh [--purge-incus]"
[ " $fail " -eq 0 ]