The loop_detect function needs improvement

[jiazhou-garland]

The loop_detect function needs improvement:

A corner case encountered during eval occurs where the agent continuously outputs the wait action, but because the wait time varies each time, loop_detect fails to stop it.

It's suggested to pop the value from the action string.

[Purewhiter]

Thanks for pointing this out.

This is indeed a real problem in the current loop detection logic. The existing implementation can miss some behaviorally repetitive actions.

We will improve this in a future update so loop detection can handle these corner cases more robustly.