Repository navigation
[Bug]: on Windows, killing a terminal leaves the process alive for five extra seconds #22
Description
Product or interface: Desktop app
Version: node-pty 1.1.0, Windows 11 x64 (10.0.26100)
Platform: Windows
OS version and architecture: Windows 11 x64 (10.0.26100)
Issue area: Tool call / permission / sandbox
What happens
On Windows, killing a terminal leaves the process alive for about five seconds longer than it has any work to do.
node-pty's Windows ConPTY kill path forks a helper (conpty_console_list_agent) to enumerate the processes attached to the shell's console, so that a killed terminal cannot leave detached processes behind. A shell hosted by ConPTY has a pseudo-console, not a console, so the helper's AttachConsole always fails there. The helper dies without answering, and node-pty falls back to a five-second timer to resolve the promise.
That timer lives in the process that killed the terminal. Every kill therefore leaves that process with five seconds of pending work it did not ask for.
Measured
Killing five powershell.exe terminals on this host:
kill -> exit events finished at 720 ms
the process was still alive at 5740 ms
one "Error: AttachConsole failed" per terminal
The AttachConsole failed lines are also what shows up in the CI logs of the native PTY acceptance case, which is what made its shutdown hook miss its deadline. That case is fixed (#17) by having the host exit once it has reported; the drag itself is still there.
What it costs
- The desktop app takes about five extra seconds to quit after the WebUI service closes with a terminal open.
- Any process that opens a terminal and then exits pays the same five seconds.
What was tried
node-pty's useConptyDll option skips the helper entirely (the shipped prebuilds/win32-x64/conpty/conpty.dll is present). Measured on this host it is worse: the process did not exit at all.
What would help
Either:
- a node-pty change that unrefs the fallback timer, or that attaches to the pseudo-console instead of a console; or
- a way for the host to stop waiting on that promise — the kill itself is synchronous, only the promise resolution is delayed.
The repository cannot fix this from its own side: the timer is internal to node-pty and the handle is not reachable from outside.
Activity
antianqi commentedon Nov 10, 2026
PR #17 (merged as c32bfcd) now works around this in the acceptance host: the host exits as soon as it has reported, so the acceptance case no longer pays the five seconds. The case went from 13-15s to about 8.5s, and it now fails loudly if the host has not exited within two seconds of reporting.
The root cause is still in node-pty and still affects every other caller on Windows. node-pty 1.1.0 forks
conpty_console_list_agenton every kill to enumerate the processes attached to the shell's console. A shell hosted by ConPTY has a pseudo-console rather than a console, so the helper'sAttachConsolealways fails there; the helper dies without answering, and node-pty resolves the promise from a five-second fallback timer that lives in the process that killed the terminal. Measured on a Windows host: killing five terminals finished in 720ms, but the process stayed alive for 5.74s, and each terminal printedError: AttachConsole failed.useConptyDll: trueskips the helper entirely and was measured too: the process did not exit at all, so it is not a workaround.So this issue is about node-pty itself: the five-second fallback should not be the only path, and the helper failing should be surfaced rather than silently absorbed.