We want to improve the UX of Lightning's AI assistant so that code suggested by the model can be automatically applied to user's code.

We will achieve this by:

• Returning a key like suggested_expression in the final payload
• Returning a key like suggested_diff in the final payload
• Have the model return a diff and explanation, rather than inline code snippets in the chat
• Apply the diff suggested to the model to the original text and send it back to the client
• Also include the raw diff that goes back to the client

Additionally, we will impose a limit (5000 characters?) with a default in the code and an override in an env var. This constraints the size of the expression that we will send to the model.

We should aim to make changes which break Lightning as little as possible.

Note that Lightning may send a partial expression (ie just the selected code) rather than the whole job. This is fine. Apollo won't know and should assume that what it is sent is the whole expression.