> Hence on top of command line applications. The power and flexibility are still there for people who want them, but so are the GUIs.
You don't want the GUI to be "on top of" the command line. A CLI makes for a terrible piece of code to integrate with!
What you really want is for both of them to share a common library. The CLI should be a client of the library just like the GUI. Usually the CLI and the library will be developed in tandem (so you have an endpoint to invoke a new library feature) but keeping them separate is a good idea.
A typical CLI might make a terrible piece of code to integrate with. I wonder if you can design CLIs that are better to integrate with, though. To some extent, this is what Git's 'plumbing' layer tries to be, although i don't know how successful it is.
One advantage of building on top of the CLI rather than via a library is that it enforces the rule that any functionality available in the GUI is also available in the CLI. It also means that any functionality added to the CLI is also immediately available to the GUI. With the library approach, it's easier for one to get ahead of the other.
Mercurial has an interesting take on this - the command server:
This lets a process spawn a sort of daemon version of Mercurial, with which it can communicate via its standard input and output. Communication is a machine-friendly framing protocol which allows the parent to execute Mercurial commands. These are the same commands as you would use on the command line (interpreted by the same dispatcher, i believe, so really exactly the same). So, it's good code to integrate with, but exactly as powerful as the CLI.
The one real let-down is that the output is, AFAIK, the same as Mercurial's interactive output, which means the parent has to parse it with various command-specific regexps. If you were following this approach from the beginning, you would probably structure your CLI so that commands emit output in some structured form (eg a tree of named, typed values, or a table of named, typed columns), which is then rendered to the console somehow. You could then return a machine-friendly serialisation of that in command server mode.
You don't want the GUI to be "on top of" the command line. A CLI makes for a terrible piece of code to integrate with!
What you really want is for both of them to share a common library. The CLI should be a client of the library just like the GUI. Usually the CLI and the library will be developed in tandem (so you have an endpoint to invoke a new library feature) but keeping them separate is a good idea.