Changing model while agent is working breaks

[zfoong]

Title

The expected behaviour should be all task sessions recreate new session cache and use the new model immediately in the next LLM call.