Can anyone explain how this is possible?

  Developers can update Claude’s instructions mid-task without breaking the prompt cache or routing the update through a user turn. This can be used in a given harness to update permissions, token budgets, or environment context as an agent runs.

Perhaps they trained it with a new special system instruction token that is specifically trained to produce the same result as changing the system prompt, but is inserted into the prompt mid-conversation?

The commands they list are app management, not part of LLM context. It's a bugfix for a needlessly delayed UI, not a model capability.