Feature Request: Improve management of mode transitions

[nissa-seru]

IIUC, current functionality is that when the user switches the mode of a model mid-task (or approves a model's request to switch modes), the model's system prompt changes accordingly but the entire current context is retained.

I strongly recommend garbage-collecting context when this occurs, as cache is broken. See #544 for recommended augmentations to the current garbage-collection functionality, but simply triggering the current functionality would be a medium win by itself. Tailored logic for this case may be incrementally valuable as well - ie it may be superior for the model to auto-generate a summary of the implementation plan and then start a new thread or more aggressively garbage-collect (ie more than half the message history) to avoid garbage-collection occurring relatively briefly after starting implementation.

There is a second issue here, though, that is likely more critical. When a task is the first message in a thread, it is protected from the context-garbage-collection logic. However, when a mode switch occurs mid-thread, that typically indicates that the user+model have crafted an implementation plan, and then the user is switching the model's mode for execution. In this case, the implementation plan is not protected from garbage-collection (!!). Hence, once the implementation plan trickles upwards into the first-half of message pairs (which is garbage-collected), it is presently removed, and the model is left to continue implementation without the benefit of the plan. IME, this results in the model getting stuck in a very narrow debug loop.