This sort of problem is why I built a stop supervisor plugin for Claude code and Codex that is automatically called whenever the agent tries to stop working. It checks to see whether there are still things to do that you had authorized. If so, it tells the agent that the stop is not authorized and to continue, and what it needs to work on.
The stop supervisor, stopContinue, can be separately enabled as part of reviewer-agent (https://github.com/rladr-org/reviewer-agent). You give it a prompt and optionally directives to control when it allows a stop and when it doesn’t. It has access to everything you’ve said in the conversation, so it can tell if you’ve authorized something that is still to do.
(It will allow the agent to stop if there are background processes that will wake it up when they finish. And it also can re-restart an agent if there are API failures.)
The stop supervisor, stopContinue, can be separately enabled as part of reviewer-agent (https://github.com/rladr-org/reviewer-agent). You give it a prompt and optionally directives to control when it allows a stop and when it doesn’t. It has access to everything you’ve said in the conversation, so it can tell if you’ve authorized something that is still to do.
(It will allow the agent to stop if there are background processes that will wake it up when they finish. And it also can re-restart an agent if there are API failures.)