long-running-jobs

How an agent should wait for work that outlives a single turn.

The principle

Polling is not the problem. Polling with the model inside the loop is.

until squeue -h -j "$JID" | grep -q .; do sleep 60; done   # 8 hours, ~0 tokens

CronCreate("check training") -> wake -> squeue -> "still epoch 21" -> reschedule
                                                  # 1 full context read PER TICK

Both poll. Only one charges you a context window every 60 seconds. When a session is carrying 400k+ tokens, a scheduled heartbeat is the single most expensive way to learn that nothing has changed.