Editing Openai/694057b6-101c-8007-9a65-c40578c7252d (section)

=== User: I think a repeated line detector would be enough beyond the token limit. ===
I think a repeated line detector would be enough beyond the token limit. It can be very simple, unlikely to have false positives, and can early-kill the worst derangements.

Another idea I'm considering is having "tools" for the ending: not meant for the normal tool round trip, but as indicators or requests for higher level decision models so it can easily and clearly indicate certain kinds of issues without needing the other model to read and understand all of the output.