Hmm. The article doesn't really describe what it's like but I skimmed the SLIME manual. So you basically point at the code you want to run, and this isn't necessarily how you would run it in production. It looks like SLIME mode requires you to install an interpreter hook in the startup code for the process that you want to debug, so you can run commands that way?

This seems like it would make the most sense for event-driven code. Evaluating a script is another kind of event. Many systems have a way to execute commands.

I'd guess this style of development encourages people to structure their systems as essentially command interpreters, even if the commands normally come from a UI.

The interesting bit will be deciding how the data structures that persist between commands are implemented. For example in server-side web apps, the shared state is often a database, so state is external to the process being debugged. The architecture limits what you can do directly, but you can try out SQL queries.