{ writeShellScriptBin
, coreutils
, findutils
, gnugrep
, iproute2
, systemd
}:

# For user services that only make sense while online (jellyfin-mpv-shim,
# nextcloud-client, ...). They keep a persistent socket to their server; when
# the machine drops off the network that socket dies, but the process stays
# alive and the unit stays "active" — running yet deaf, the "up but not
# working, restart it by hand" case. systemd can't see this: nothing crashes
# and the programs speak no health protocol, so we probe the real thing — is
# the MainPID still holding a live (established, non-loopback) connection?
#
#   wait-online UNIT  block until a real address exists; used as ExecStartPre
#                     so UNIT only starts once online. Also stamps the start
#                     time so `check` leaves a freshly (re)started UNIT alone.
#   check UNIT...     one tick, run by a 60s timer: restart every UNIT that is
#                     active but not connected, at most once per 5 minutes.
#
# Wired up by the services.network-watchdog home-manager module.

writeShellScriptBin "network-watchdog" ''
  set -u
  stampdir="''${XDG_RUNTIME_DIR:-/run/user/$(${coreutils}/bin/id -u)}/network-watchdog"
  ${coreutils}/bin/mkdir -p "$stampdir"

  usage() { echo "usage: network-watchdog wait-online UNIT | check UNIT..." >&2; exit 64; }

  # A global-scope address on a non-loopback interface — true for a LAN IP too,
  # so this needs no server URL and works whether the server is local or remote.
  network_up() { ${iproute2}/bin/ip -o addr show scope global up 2>/dev/null | ${gnugrep}/bin/grep -q .; }

  # Healthy: MainPID owns an established socket to something other than loopback.
  connected() {
    pid="$(${systemd}/bin/systemctl --user show -p MainPID --value "$1" 2>/dev/null)"
    [ "''${pid:-0}" != 0 ] && ${iproute2}/bin/ss -tnpH state established 2>/dev/null \
      | ${gnugrep}/bin/grep "pid=$pid," \
      | ${gnugrep}/bin/grep -qv -e '127\.0\.0\.1' -e '\[::1\]'
  }

  # Stamped within the last 5 minutes: still connecting, or already restarted.
  recently_started() { [ -n "$(${findutils}/bin/find "$stampdir/$1" -mmin -5 2>/dev/null)" ]; }

  [ $# -ge 1 ] || usage
  cmd="$1"; shift
  case "$cmd" in
    wait-online)
      [ $# -eq 1 ] || usage
      while ! network_up; do ${coreutils}/bin/sleep 5; done
      : > "$stampdir/$1"
      ;;
    check)
      # Nothing can be connected while offline, and restarting wouldn't help.
      network_up || exit 0
      for unit; do
        # Not "active" (inactive, failed, or still in wait-online): systemd's job, not ours.
        ${systemd}/bin/systemctl --user is-active --quiet "$unit" || continue
        recently_started "$unit" && continue
        connected "$unit" && continue
        echo "network-watchdog: $unit is online but has no live connection -> restarting"
        ${systemd}/bin/systemctl --user restart "$unit"  # its wait-online refreshes the stamp
      done
      ;;
    *) usage ;;
  esac
''