Repository navigation
attach restarts a dead session without consent when stdin is not a TTY (EOF defaults to yes) #214
Description
@myobie, Compounding Manager reports a pre-existing behavior from a live smoke test of #209 (evidence: schickling/dotfiles#8391 comment 6620845281).
Observed: pty attach <name> on a session that pty classifies as dead (here: a record whose daemonStartToken does not match the live PID) prints Restart? [Y/n]. When stdin is not a terminal or is at EOF, the answer defaults to yes. attach then removes the old record, spawns a new daemon that reruns the stored command, and attaches to it. --force behaves the same. Reproduced with 74e4dbb and with rust.5 (65bf24d), so #209 did not introduce it.
PTY_ROOT="$ROOT" pty attach negative < /dev/null
# Session "negative" exited with code unknown.
# Restart? [Y/n] Session "negative" restarted.Code path at 74e4dbb: attach.rs:189,227-264 and ask.rs:13-30.
Why it matters: agents call attach non-interactively. A non-interactive caller cannot consent, yet the stored command runs again. For mutating commands this repeats side effects (same concern as #204).
Question: should a prompt with no answer (no TTY or EOF) default to "no" and exit non-zero, so that a restart needs an explicit y or a flag such as attach -r?
Safety part passed: the unrelated process that reused the PID was never attached to or signalled.
Posted on behalf of @schickling
| field | value |
|---|---|
agent_identity |
agent/compoundingtech/compounding-manager |
session |
01a1d992 |
agent_persona |
generalist |
agent_supervisor |
unavailable |
agent_tool |
OMP |
agent_tool_version |
18.12.5 |
agent_runtime |
OMP 18.12.5 |
agent_model |
anthropic/claude-opus-5-5 |
worktree |
unknown |
Activity
myobie commentedon Nov 4, 2026
I measured this at main
ddfcf093, with no terminal.What the code does today
attachon a dead session prints the last lines and asksRestart? [Y/n](attach.rs#L231).ask.rsreads one line with no terminal check, and only a lonendeclines. EOF reads as an empty answer, so</dev/nullrestarts (ask.rs#L13-L31).--forceonly allows attaching from inside another session. It never answered the prompt, so it restarts the same way.-rrestarts without asking.--no-restartrefuses with exit 1 before the stored command is considered, and it is the safe flag for callers today.A correction.
ask.rssays a closed stdin proceeds "exactly like Node", and I repeated that on #204. It's wrong for EOF. I built the pinned Node reference (f005432a) and ran it with no terminal, on Node 22 and on 24. At EOF itsreadlinequestion never resolves, and the process exits 0 without restarting. It does restart on any piped line butn, an empty line included.ddfcf093</dev/nullyoryesn-r,</dev/null--force,</dev/nullThe change, decided and in #215.
yoryesrestarts, and EOF never does.attachexits 1 and keeps the exited record. It saysAnswer y, or pass -r to restart it without asking.nexits 0,-rrestarts without asking, and--forceanswers nothing.Decision 0020 records the departure from Node. A script that relied on a non-interactive restart must pass
-r.The other prompt with this shape is
pty restart'sKill and restart? [Y/n]. With</dev/nullit stops a running session and starts it again. Node exits 0 and leaves the session running. #215 leaves it unchanged; it's listed there for a separate fix.Merge pull request #215 from compoundingtech/agent/attach-unanswered-…