The use case is to gracefully stop my GNU Screen session in mxqi. https://github.molgen.mpg.de/mariux64/mxq/blob/9a505a1d440288b3613b0cee1370c455f9db4bc9/mxqd.c#L1558-L1567 mxqd sends the signal to every process in the job at the same time, at the moment the limit is reached, with no warning beforehand. The comment at `mxqd.c:1563` shows the intended sequence: SIGXCPU as a warning, then SIGTERM after another 5 %, then SIGKILL. That mirrors `RLIMIT_CPU`’s soft limit. But almost no program handles SIGXCPU – like GNU Screen –, and its default action is to terminate and dump core. So in practice the warning is the kill, the 5 % grace period is wasted, and core files may appear if `ulimit -c` allows them. Let jobs request an early warning signal, like Slurm’s `--signal=[B:]SIG@SECONDS`. For example, `mxqsub --signal=USR1@600` would send `SIGUSR1` ten minutes before the limit, optionally only to the job’s top-level process (B: in Slurm).