Claude on Question 14, Sep 19

For a well-behaved exit, the idiom is a trap plus killing the child process group — not individual PIDs, since grandchildren are what actually leak. For the "parent got SIGKILLed" case, traps can't help you at all and you need kernel-level help (--pdeathsig, PID namespaces, or cgroups).

The common case: trap + process group

The key detail people miss: a non-interactive bash script is not a process-group leader, it inherits its caller's group. So kill -- -$$ can nuke your terminal or CI runner. Turn on job control so each background job gets its own group:

#!/usr/bin/env bash
set -euo pipefail
set -m                      # job control: each bg job becomes a pgroup leader

children=()

spawn() {
  "$@" &
  children+=("$!")          # with set -m, $! is also the PGID
}

cleanup() {
  trap - EXIT INT TERM      # don't re-enter
  local pgid
  for pgid in "${children[@]}"; do
    kill -TERM "-$pgid" 2>/dev/null || true
  done
  # give them a moment, then escalate
  sleep 0.5
  for pgid in "${children[@]}"; do
    kill -KILL "-$pgid" 2>/dev/null || true
  done
  wait 2>/dev/null || true  # reap
}
trap cleanup EXIT INT TERM

spawn ./some-server
spawn ./some-worker
wait

The - prefix on the PID is what makes kill target the whole group. Recording the PGID up front matters because once the parent dies the orphans are reparented to init (or a subreaper), so pkill -P $$ won't find them anymore.

If you only ever have direct children and don't care about trees, the one-liner version is fine:

trap 'kill $(jobs -p) 2>/dev/null' EXIT

When the parent is SIGKILLed

No trap runs, so you need the kernel to do the cleanup. Three options, roughly in order of how invasive they are:

setpriv --pdeathsig (util-linux ≥ 2.33) — wraps prctl(PR_SET_PDEATHSIG). Each child gets signalled when its parent dies:

setpriv --pdeathsig=TERM -- ./some-server &

Caveat: it fires when the direct parent's thread exits, so it doesn't propagate down a tree on its own, and it's lost across an intermediate process that doesn't re-arm it.

PID namespace — unshare --pid --fork --kill-child makes your script PID 1 of a new namespace; when it dies, the kernel kills everything inside:

unshare --pid --fork --mount-proc --kill-child -- ./my-script.sh

Needs privileges or CONFIG_USER_NS / unprivileged userns enabled.

cgroup/scope — most robust on systemd hosts, and what service managers actually use:

systemd-run --user --scope --collect -- ./my-script.sh

Everything lands in one cgroup that gets torn down as a unit.

Two other notes

• •If you want to reap orphans rather than kill them (e.g. you're a supervisor and want to keep collecting exit statuses), that's PR_SET_CHILD_SUBREAPER, not available from bash directly — use a real init like tini or dumb-init as PID 1.
• •shopt -s huponexit only applies to interactive login shells; it won't do anything in a script.

For most scripts, the set -m + trap version above plus setpriv --pdeathsig on long-running children covers both graceful and abrupt death without needing privileges.