Just a wild guess but could a race condition between a subtask marking itself complete and your “supervisor” process that checks if no subtasks are incomplete.

So when subtask 3 for example marks itself complete if your “supervisor” checks before your subtask 4 is added it would see all subtasks complete. A bandaid could be only marking a subtask complete after the next subtask is inserted.

Sounds like you’re using a DB to track the jobs. Is that because your workers are distributed?