tui: Kitty graphics probe (Gi=31337) leaks into the shell prompt after exiting V2 #55641

Activity

1. opencode-agent commented on Oct 15, 2026

   

   opencode-agentBot

   on Oct 15, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the detailed report. I can reproduce this on 2.0.29 in qTerminal (TERM=xterm-256color, bash, no tmux). After exiting the TUI, …73ppppppGi=31337,s=1,v=1,a=q,t=d,f=24;AAAA is left on the screen. The recording and screenshot below show it, and the script below reproduces it under Xvfb.

   I think the cause is a bit different from #17081, though. A raw capture of the terminal output shows OpenCode sends its capability probes on the normal screen before it switches to the alternate screen:

   • an XTGETTCAP query (ESC P +q4d73 ESC \)
   • six DECRQM mode queries ending in $p
   • the Kitty graphics query (ESC _ Gi=31337…;AAAA ESC \)

   qTerminal doesn't understand DCS/APC sequences, so it draws their contents as text: +q4d73, then a p from each mode query, then Gi=31337…. That text is still on the normal screen when OpenCode exits.

   In my test this was only on the screen. Typing a command right after exiting worked normally, so the shell's input line held no stray bytes, and I couldn't reproduce the Ctrl+D problem. If Ctrl+D is still blocked for you, it would help to know what's on the line when you press it. For example, does echo hi run cleanly right after exiting?

   I couldn't run the latest dev (4397ee4) from source here. But the latest release is still 2.0.29, and the terminal library that sends these probes (@opentui 0.5.17) is the same version on dev, so it most likely still happens there.

   recording.mp4

   

   Reproduction script (qterm-repro.sh)

   #!/bin/bash
   # Reproduces #55641: launch opencode inside qTerminal (under Xvfb), exit, screenshot the shell prompt.
   # Usage: OC_CMD="opencode" ./qterm-repro.sh   (needs qterminal, Xvfb, xdotool, imagemagick; ffmpeg optional)
   set -u
   OC_CMD=${OC_CMD:-opencode}
   OUT=${OUT:-/workspace/media}
   mkdir -p "$OUT" /workspace/repro/proj
   export DISPLAY=:99
   Xvfb :99 -screen 0 1280x800x24 >/dev/null 2>&1 & XV=$!
   sleep 2
   REC=""
   if command -v ffmpeg >/dev/null; then
     ffmpeg -loglevel error -y -f x11grab -video_size 1280x800 -framerate 10 -i :99 -pix_fmt yuv420p "$OUT/qterminal-repro.mp4" & REC=$!
   fi
   cd /workspace/repro/proj
   env -u TERM_PROGRAM -u TMUX qterminal >/dev/null 2>&1 & QT=$!
   sleep 4
   WID=$(xdotool search --pid $QT | tail -1)
   xdotool windowactivate --sync "$WID" 2>/dev/null; xdotool windowsize "$WID" 1200 760 windowmove "$WID" 0 0
   sleep 1
   xdotool type --delay 50 "echo TERM=\$TERM; $OC_CMD"; xdotool key Return
   sleep 12
   import -window root "$OUT/qterminal-in-tui.png"
   xdotool key ctrl+c; sleep 1; xdotool key ctrl+c
   sleep 4
   xdotool type --delay 50 "echo READLINE_CLEAN"; xdotool key Return; sleep 2
   import -window root "$OUT/qterminal-after-exit.png"
   sleep 1
   [ -n "$REC" ] && kill -INT $REC && sleep 2
   kill $QT $XV 2>/dev/null

   Run it: bash qterm-repro.sh
2. opencode-agent

   added

   reproduced

   on Oct 15, 2026
3. dersenok commented on Oct 15, 2026

   

   dersenok

   on Oct 15, 2026

   Author

   More actions

   Reporter here — thanks for the detailed repro. Confirming from my side, with one difference from your run:

   Ctrl+D does not work right after exit, and unlike your test there is a stray character sitting on the readline line. Pressing Enter (or Ctrl+L) clears it, and then Ctrl+D works again. So on my machine both halves are present:

   1. the DCS/APC probe text drawn on the normal screen (exactly as you describe: +q4d73 + six ps + Gi=31337,…), and
   2. at least one stray byte left in the tty input buffer — the Stale stdin bytes leak to shell after exit on POSIX (space on prompt, blocks ctrl+d) #17081 symptom — which sits on the line and blocks EOF.

   So it looks like V2 may have regressed the #17081 tcflush fix while the screen-scrape half (probes emitted before entering the alt screen) is separate.

   Env: qTerminal 2.1.0, TERM=xterm-256color, COLORTERM=truecolor, /bin/bash, no tmux, opencode 2.0.29.

   Happy to run a targeted diagnostic if useful — e.g. dump the readline line as octal right after exit, before touching the keyboard, to identify which byte leaks into the input buffer. Just tell me what you'd like captured.
4. Finn763 commented on Oct 21, 2026

   

   Finn763

   on Oct 21, 2026

   Contributor

   More actions

   I'd like to work on this — the Kitty graphics probe leaves its response fragment unread at exit; I'll clear it in the TUI shutdown path.