mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-06 05:17:15 +00:00
kshitijk4poor
e883b5a4d5
Review (nit 1): `sleep` is not a dash builtin, so with an unusable PATH (`sleep: not found`) the supervisor's three one-second waits returned at once and the KILL followed the TERM at +0 s, cutting short a child's TERM cleanup. The wait now runs `command -p sleep 1` (the shell's default PATH), falling back to a plain `sleep 1` on a shell without `command -p`. Test: with PATH=/nonexistent a TERM-ignoring group is still alive 1.5 s after the EOF and is then KILLed; fails with the plain `sleep`.
57 lines
2.8 KiB
Bash
Executable File
57 lines
2.8 KiB
Bash
Executable File
#!/bin/sh
|
|
# Hermes process supervisor: one per Hermes process, reaping the process groups it owns once it dies.
|
|
#
|
|
# Usage: sh process_supervisor.sh <parent-pgid>, spawned with start_new_session=True and stdin a
|
|
# pipe whose only write end Hermes holds (hermes_cli/process_identity.py, "owner-death supervisor").
|
|
# stdin carries one command per line:
|
|
# register <pgid> reap this process group when Hermes dies
|
|
# unregister <pgid> stop reaping it (Hermes stopped that child itself)
|
|
# EOF means Hermes is gone, by any cause including SIGKILL and the OOM killer: SIGTERM every group
|
|
# still registered, wait up to 3 s, SIGKILL the survivors, exit. A clean exit is the same EOF, so a
|
|
# child Hermes never unregistered dies with it either way.
|
|
#
|
|
# POSIX sh, not Python: Hermes keeps this process resident for its whole life, and an idle sh costs
|
|
# about 0.5 MB USS where an idle interpreter costs about 8 MB. It must also never import Hermes:
|
|
# it runs after Hermes may already be dead.
|
|
#
|
|
# Malformed lines are skipped: a corrupt line must not cost the reaping guarantee for every other
|
|
# group. Groups that are never a valid target (pgid <= 1, our own, Hermes's own) are refused at
|
|
# registration, so a caller bug costs one unreaped child, never the Hermes process tree.
|
|
# Residual risk: groups are reaped by number, so a registered group that exits on its own can have
|
|
# its pgid recycled before Hermes prunes it; Hermes prunes empty groups on every registration change.
|
|
|
|
trap '' INT HUP # a dying parent's SIGINT/SIGHUP must not preempt the reap: EOF is the only signal
|
|
parent=$1
|
|
# Every register line compares against it: a bad value would silently refuse every registration.
|
|
case $parent in ''|*[!0-9]*) echo "hermes process supervisor: bad parent pgid '$parent'; refusing to run" >&2; exit 2 ;; esac
|
|
# Run inside the parent's group, a killpg of the parent would take us down before we reap.
|
|
kill -0 -$$ 2>/dev/null || {
|
|
echo "hermes process supervisor: not a process-group leader; refusing to run" >&2
|
|
exit 2
|
|
}
|
|
reg=' '
|
|
while IFS=' ' read -r verb pg rest; do
|
|
case $pg in ''|*[!0-9]*) continue ;; esac
|
|
[ -z "$rest" ] && [ ${#pg} -le 10 ] || continue
|
|
case $verb in
|
|
register)
|
|
[ "$pg" -gt 1 ] && [ "$pg" -ne $$ ] && [ "$pg" -ne "$parent" ] || continue
|
|
case $reg in *" $pg "*) ;; *) reg="$reg$pg " ;; esac ;;
|
|
unregister)
|
|
keep=' '
|
|
for g in $reg; do [ "$g" = "$pg" ] || keep="$keep$g "; done
|
|
reg=$keep ;;
|
|
esac
|
|
done
|
|
for g in $reg; do kill -TERM -"$g" 2>/dev/null; done
|
|
for _ in 1 2 3; do
|
|
live=''
|
|
for g in $reg; do kill -0 -"$g" 2>/dev/null && live="$live $g"; done
|
|
reg=$live
|
|
[ -n "$reg" ] || exit 0
|
|
# The default PATH: a broken PATH must not turn the TERM grace into an instant KILL.
|
|
command -p sleep 1 2>/dev/null || sleep 1
|
|
done
|
|
for g in $reg; do kill -KILL -"$g" 2>/dev/null; done
|
|
exit 0
|