diff --git a/bin/box b/bin/box
index 24c24bb..6511657 100755
--- a/bin/box
+++ b/bin/box
@@ -472,7 +472,7 @@ wait_agent() {
   local n="$1" i
   echo "box: waiting for instance agent..."
   for i in $(seq 1 90); do
-    if incus exec "$n" -- true >/dev/null 2>&1; then return; fi
+    if incus exec "$n" -- true </dev/null >/dev/null 2>&1; then return; fi
     [ "$i" -eq 90 ] && die "instance agent never came up (incus console $n to inspect)"
     sleep 2
   done
@@ -500,7 +500,7 @@ reset_identity() {
     systemd-machine-id-setup >/dev/null 2>&1 || dbus-uuidgen > /etc/machine-id
     ln -sf /etc/machine-id /var/lib/dbus/machine-id
     test -s /etc/machine-id
-  ' || die "could not reset the clone's machine-id"
+  ' </dev/null || die "could not reset the clone's machine-id"
   # The new id only takes effect at boot. Ask nicely, then insist — a clone that
   # keeps its source's DHCP lease is worse than an unclean stop of a box that
   # booted 30 seconds ago.
@@ -592,15 +592,20 @@ cmd_new() {
     # A failed cloud-init used to print a screen of dots and the word "error",
     # with nothing to act on — the box's own log holds the reason, and nobody
     # was told it existed. Show it, and leave the box up to inspect.
-    if ! incus exec "$instance" -- cloud-init status --wait; then
+    # Every non-interactive exec pins stdin. With a TTY on stdin, 'incus exec'
+    # goes interactive — and when box's own output is redirected (a script, the
+    # drill), the session can wedge open after the remote command has exited,
+    # blocking forever on a websocket that will never close. Caught live: a
+    # mint stuck at 'status: done'. Only shell/exec/tmux may own the terminal.
+    if ! incus exec "$instance" -- cloud-init status --wait </dev/null; then
       echo >&2
       echo "box: cloud-init FAILED in $name. What it says:" >&2
-      incus exec "$instance" -- cloud-init status --long 2>&1 | sed 's/^/  /' >&2
+      incus exec "$instance" -- cloud-init status --long </dev/null 2>&1 | sed 's/^/  /' >&2
       echo >&2
       echo "box: the errors, from the box's log:" >&2
       incus exec "$instance" -- sh -c \
         "grep -iE '^(E:|Err:)|Temporary failure|Could not resolve|Unable to fetch' /var/log/cloud-init-output.log | tail -8" \
-        2>/dev/null | sed 's/^/  /' >&2
+        </dev/null 2>/dev/null | sed 's/^/  /' >&2
       echo >&2
       echo "box: '$name' is still up — inspect it, then delete it:" >&2
       echo "  box incus $name -- exec {} -- tail -50 /var/log/cloud-init-output.log" >&2
diff --git a/drill/wipe.sh b/drill/wipe.sh
index d16e6d5..e368f5a 100644
--- a/drill/wipe.sh
+++ b/drill/wipe.sh
@@ -3,11 +3,13 @@
 # pre-0.4.0 claudebox, so the next drill run starts from a truly bare host and
 # its verdict means something.
 #
-#   bash drill/wipe.sh                   # asks first
+#   bash drill/wipe.sh                   # asks first; KEEPS cached images (an
+#                                        # image is upstream's artifact — wiping
+#                                        # it buys nothing but a re-download)
 #   bash drill/wipe.sh --yes             # no prompt
-#   bash drill/wipe.sh --purge-storage   # also delete cached images AND the
-#                                        # 'default' storage pool, so setup-host
-#                                        # exercises its pool bootstrap (#29)
+#   bash drill/wipe.sh --purge-storage   # also delete the 'default' storage
+#                                        # pool (and the images inside it), so
+#                                        # setup-host exercises its bootstrap (#29)
 #
 # What teardown-host.sh does NOT cover, this does: instances the drill names
 # but never tagged, instances of either tag generation, cached images, and
@@ -40,8 +42,8 @@ This wipes EVERY trace of box/claudebox from this host ($(hostname)):
   · networks boxnet + claudenet, ACLs box-isolate + claude-isolate
   · profiles box-net + claude-dev
   · firewall units, scripts and nft tables of BOTH name generations
-  · every cached Incus image (the next mint re-downloads)
-$( [ "$PURGE_STORAGE" = 1 ] && echo "  · the 'default' storage pool (--purge-storage)" )
+$( [ "$PURGE_STORAGE" = 1 ] && echo "  · the 'default' storage pool AND its cached images (--purge-storage)" \
+                            || echo "  · (cached images are KEPT — the next mint stays fast; --purge-storage removes them with the pool)" )
 Uncommitted work inside any box is LOST. Only do this on a drill host.
 EOF
   [ -t 0 ] || { echo "wipe: no TTY to confirm on — pass --yes if you mean it." >&2; exit 2; }
@@ -79,13 +81,16 @@ if command -v incus >/dev/null; then
     incus network acl delete "$acl" >/dev/null 2>&1 && say "deleted ACL $acl"
   done
 
-  # --- cached images: the pool's other tenants -------------------------------
-  for f in $(incus image list -f csv -c f 2>/dev/null); do
-    incus image delete "$f" >/dev/null 2>&1 && say "deleted image $f"
-  done
-
-  # --- the pool itself (opt-in): lets setup-host's bootstrap run for real ----
+  # --- cached images: NOT wiped by default -----------------------------------
+  # An image is upstream's artifact, content-addressed by fingerprint — not a
+  # drill artifact. Deleting it buys zero cleanliness and costs the next mint
+  # a full re-download. It only goes when the pool it lives in goes.
   if [ "$PURGE_STORAGE" = 1 ]; then
+    # --- the pool (opt-in): lets setup-host's bootstrap run for real ---------
+    # Images live in the pool and block its deletion — they go first.
+    for f in $(incus image list -f csv -c f 2>/dev/null); do
+      incus image delete "$f" >/dev/null 2>&1 && say "deleted image $f"
+    done
     incus profile device remove default root >/dev/null 2>&1 && say "removed default profile's root device"
     if incus storage delete default >/dev/null 2>&1; then
       say "deleted storage pool 'default' — setup-host will rebuild it (btrfs where it can)"