ACP-friendly `think` and `do` syntax

[dherman]

The think and do syntaxes should be designed to integrate cleanly with the ACP REPL experience.

Desired REPL experiences

• Top level code mode:

{ print("hi") }

• Top level shell mode:

$ ls

• Top level prompt mode:

How much wood could a woodchuck chuck if a woodchuck could chuck wood?

• Standard stream or API for producing thought chunks
• Rename think operator to chat?
• Remove do operator in favor of just curlies? Does that create too much of an ambiguity with markdown?

[dherman]

Let's stick with think, not chat. There's just one LLM per context, so it's part of the current agent. So the mental model should be that using the LLM is internal to the agent, i.e. a part of its internal thought process. Thought chunks are communicating outside of the agent what it's currently thinking about—a kind of agent log.