6433ab32

test: a down host says why on the row, and nothing on the wall

a73x 2026-08-30 10:58

Commit message

test: a down host says why on the row, and nothing on the wall

Two legs at the end of the hosts group, the bug's own shape: a hosts line
naming a box that is off, polled through an ssh stand-in that dies the
way an unplugged machine does.

The wall leg asserts on the client's own fd 2 — which on a terminal is
the alternate screen it is painting — and on the capture as a COUNT,
because mux now puts that sentence on the screen itself, inside the
picker row. Every copy must be one the row painted.

The entry-dial leg is the other half: `narrate` is its alone, so it
relays ssh's bytes as they come, and its last line quotes the same
sentence instead of an error name only mux understands.

The two handoff legs that read the remote's progress off the user's
stderr keep passing unchanged — they are entry dials, and their comments
now say relayed where they said inherited.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017wi2HnuF1EK8HgViU11YLV

test/e2e.sh

Old		New
@@ -170,8 +170,8 @@ done
170	# one of those and adds a convergence point would be pinning a fact every	170	# one of those and adds a convergence point would be pinning a fact every
171	# leg above already establishes.	171	# leg above already establishes.
172		172
173	[ "$OK_COUNT" = "91" ] \|\| {	173	[ "$OK_COUNT" = "93" ] \|\| {
174	echo "e2e FAIL: $OK_COUNT scenario checkpoints ran, the pin says 91 —"	174	echo "e2e FAIL: $OK_COUNT scenario checkpoints ran, the pin says 93 —"
175	echo " a scenario was added (update the pin) or silently lost"	175	echo " a scenario was added (update the pin) or silently lost"
176	exit 1	176	exit 1
177	}	177	}

test/e2e_04_handoff.sh

test/e2e_09_hosts.sh

@@ @@ -1114,8 +1114,8 @@ NASKS_3B=$(grep -c -- '--start' "$NSHIMLOG" || true) @@
+# (4) A cold `mux HOST` entry, on the box the wall left alone. One daemon,
+# and the REMOTE said so in its own voice — `mux d endpoint --start`'s
+# progress reaches the user's terminal over the same relayed stderr ssh's
+# own diagnostics ride on, `narrate` being the entry dial's alone.
+pipe_mux "$OUT.nsa" "$OUT.nsa.err" env SHELL=/bin/sh XDG_STATE_HOME="$NSTATE" \
+    PATH="$NPATH" timeout 60 "$MUX" "nostart-a@127.0.0.1"
+pipe_send 'printf "ns-cold-%%s\\n" ok\n'
@@ @@ -1157,3 +1157,120 @@ NASKS_ALL=$(grep -c -- '--start' "$NSHIMLOG" || true) @@
+    cat "$OUT.nsastop"; exit 1; }
+nostart_empty "$NRUNA" "$NSOCKA" "mux d stop"
+ok "reading a host leaves it alone; a picker Enter and a cold mux HOST each start one daemon"
+# ---- ssh's stderr is mux's: no byte of it reaches the wall --------------
+#
+# The bug, as the user met it: a hosts line naming a box that is down. The
+# wall polls it over ssh, ssh dies with `ssh: connect to host ... No route
+# to host`, and that line landed on the wall's ALTERNATE SCREEN every few
+# seconds — over tiles and rails, from a writer no repaint of ours can
+# reach, because the ssh child's stderr was inherited.
+#
+# It is a pipe mux reads now. The bytes go nowhere on their own; the last
+# line becomes the dial's reason, and the picker row — the one place that
+# was saying a bare `unreachable` — quotes it.
+#
+# The oracle for the regression is the client's own STDERR FILE, not the
+# capture. `ptyclient` requires `--err PATH` and gives the client fd 2 as
+# that file, so an inherited ssh writes there and never onto the pty —
+# the capture cannot see the original bug at all. The file is the
+# faithful stand-in precisely because it is the same fd: on a real
+# terminal fd 2 IS the screen mux is painting, under the alternate
+# screen, which is where those bytes landed. The capture is the second
+# pin, and a different claim — that every copy of the sentence ON SCREEN
+# is one the picker row painted.
+RSHIM="${TMPDIR:-/tmp}/mux-e2e-reason-shim-$$"
+defer_rm "$RSHIM"
+RSTATE="${TMPDIR:-/tmp}/mux-e2e-reason-state-$$"
+defer_rm "$RSTATE"
+mkdir -p "$RSHIM" "$RSTATE/mux"
+RSAID='ssh: connect to host 10.255.255.1 port 22: No route to host'
+# An ssh that fails the way an unplugged box fails: one line on stderr,
+# 255, and nothing on stdout at all. 10.255.255.1 is a literal in the
+# message, never dialled — the shim never reaches the network.
+ssh_shim_head "$RSHIM/ssh"
+cat >> "$RSHIM/ssh" <<SHIM
+printf '%s\n' '$RSAID' >&2
+exit 255
+SHIM
+chmod +x "$RSHIM/ssh"
+RPATH="$RSHIM:$(cd "$(dirname "$MUX")" && pwd):$PATH"
+printf 'noroute@127.0.0.1\n' > "$RSTATE/mux/hosts"
+# The wall. One host, no sessions, so the picker opens by itself — which
+# is also the row this leg is about. The settle is for the paint, not for
+# a second cycle: a failed `.hand` poll answers `.pipe`, so `pollDelayMs`
+# stretches the next one to ten seconds and three of them see exactly one.
+# One is enough — the first cycle is where the inherited write happened.
+set +e
+SHELL=/bin/sh XDG_STATE_HOME="$RSTATE" PATH="$RPATH" timeout 60 "$PTYCLIENT" \
+    --cols 100 --rows 30 --out "$OUT.rcap" --err "$OUT.rcap.err" -- \
+    "$MUX" > "$OUT.rpc" 2>&1 <<'EOF'
+expect unreachable 25000
+settle 3000 25000
+send \x1b
+settle 500 15000
+send \x1cd
+waitexit 15000
+EOF
+RC=$?
+set -e
+[ "$RC" -eq 0 ] || {
+    echo "e2e FAIL: reason: the wall leg exited $RC:"
+    cat "$OUT.rpc"; echo "--- stderr ---"; cat "$OUT.rcap.err"; exit 1; }
+# THE regression. A wall's own stderr is the terminal it is painting —
+# under the alternate screen, over tiles and rails — and an INHERITED ssh
+# writes straight onto it. `ptyclient` gives the client a real stderr
+# FILE, so that file is where those bytes land here, and its emptiness of
+# them is the same claim the screen makes in production. Mutate
+# `spawnPipe`'s `.Pipe` back to `.Inherit` and this is what fires
+# (measured: one copy of the line, above `mux: left the wall`).
+grep -qF "$RSAID" "$OUT.rcap.err" && {
+    echo "e2e FAIL: reason: ssh's own stderr reached the wall's own fd 2,"
+    echo "          which on a terminal is the alternate screen:"
+    cat "$OUT.rcap.err"; exit 1; }
+# And on the SCREEN, as a count rather than an absence: mux itself puts
+# ssh's sentence there now, inside the picker row, so "the bytes are not
+# there" is no longer the question. Every copy in the capture must be one
+# the row painted.
+# `wc -l` on the matches and never `grep -c`, which counts LINES: a pty
+# capture is escape bytes with almost no newlines in it, so five raw
+# copies and one row would both come back as 1.
+RALL=$(grep -aoF "$RSAID" "$OUT.rcap" | wc -l)
+RROW=$(grep -aoF "unreachable: $RSAID" "$OUT.rcap" | wc -l)
+[ "$RROW" -gt 0 ] || {
+    echo "e2e FAIL: reason: the picker row never quoted ssh's line:"
+    grep -aoF 'unreachable' "$OUT.rcap" | head -3; exit 1; }
+[ "$RALL" -eq "$RROW" ] || {
+    echo "e2e FAIL: reason: ssh's own stderr reached the wall's screen —"
+    echo "          $RALL copies of its line, only $RROW of them in a row"
+    exit 1; }
+ok "a down host's ssh says why on the picker row and nothing on the wall"
+# ---- ...and the entry dial says it in mux's own failure line ------------
+#
+# The other half of the rule: `narrate` is the entry dial's alone, so this
+# is the ONE caller that relays ssh's bytes onward as they come — there is
+# no wall yet, and the user is sitting in front of the wait. The failure
+# line then quotes the same sentence, because `UnterminatedLine` names
+# what mux observed while `No route to host` names what happened.
+set +e
+XDG_STATE_HOME="$RSTATE" PATH="$RPATH" timeout 40 "$MUX" "badhost@127.0.0.1" \
+    > "$OUT.rentry" 2> "$OUT.rentry.err"
+RC=$?
+set -e
+[ "$RC" -ne 0 ] || {
+    echo "e2e FAIL: reason: a dial that never reached a daemon exited 0"; exit 1; }
+# Relayed, not swallowed: the bytes ssh wrote reached the user's stderr.
+grep -qF "$RSAID" "$OUT.rentry.err" || {
+    echo "e2e FAIL: reason: the entry dial did not relay ssh's stderr:"
+    cat "$OUT.rentry.err"; exit 1; }
+# The LAST mux line is the verdict, and it is ssh's sentence rather than
+# an error name only mux understands.
+RLAST=$(grep -a '^mux: ' "$OUT.rentry.err" | tail -1)
+[ "$RLAST" = "mux: badhost@127.0.0.1 over ssh: $RSAID" ] || {
+    echo "e2e FAIL: reason: the entry dial's failure line is"
+    echo "          '$RLAST'"
+    echo "          want 'mux: badhost@127.0.0.1 over ssh: $RSAID'"
+    cat "$OUT.rentry.err"; exit 1; }
+ok "the entry dial relays ssh's stderr and fails in ssh's own words"

Old		New
@@ -180,7 +180,8 @@ pipe_detach "cold handoff client"
180	# The daemon the REMOTE started, by the pid its own up-line reported — the	180	# The daemon the REMOTE started, by the pid its own up-line reported — the
181	# only handle this suite has on a process that is nobody's child. The line	181	# only handle this suite has on a process that is nobody's child. The line
182	# is `mux d endpoint`'s now, and it reaches this file the way ssh's own	182	# is `mux d endpoint`'s now, and it reaches this file the way ssh's own
183	# stderr does: inherited, because the entry dial is not `quiet`. Same	183	# stderr does: RELAYED, byte for byte, because the entry dial is the one
		184	# caller that `narrate`s the pipe it reads. Same
184	# accepted gap as the M13 blocks: a failure above leaves it untracked, and	185	# accepted gap as the M13 blocks: a failure above leaves it untracked, and
185	# the trap's stop-by-socket is what covers that.	186	# the trap's stop-by-socket is what covers that.
186	grep -q '^mux d endpoint: starting' "$OUT.h1.err" \|\| {	187	grep -q '^mux d endpoint: starting' "$OUT.h1.err" \|\| {
@@ -522,7 +523,7 @@ wait_for "$OUT.h5" "none-ok" 20 \|\| {
522	HT3=$(date +%s%N)	523	HT3=$(date +%s%N)
523	HMS_E=$(( (HT3 - HT2) / 1000000 ))	524	HMS_E=$(( (HT3 - HT2) / 1000000 ))
524	pipe_detach "announce-none client"	525	pipe_detach "announce-none client"
525	# The remote said why, in one line, on the stderr ssh already carries.	526	# The remote said why, in one line, on the stderr mux relays from ssh.
526	grep -q '^mux d endpoint: .*staying on ssh' "$OUT.h5.err" \|\| {	527	grep -q '^mux d endpoint: .*staying on ssh' "$OUT.h5.err" \|\| {
527	echo "e2e FAIL: announce-none said nothing about why it stayed on ssh"	528	echo "e2e FAIL: announce-none said nothing about why it stayed on ssh"
528	cat "$OUT.h5.err"; exit 1; }	529	cat "$OUT.h5.err"; exit 1; }