First of all, there were some encoding conflicts that were introduced when mixing Leo with other editors. When I go back to the project now, it doesn't compile because of the encoding errors. (It's fixable, I'm sure.)
Second, I had more classes than I needed. The secret to writing compact game code is - basically - to write few real classes and rely on plain old data and a large main loop. As it was, they were calling up and down some hierarchy, splitting pieces of the main loop into different classes, running a custom scripting language to drive AI, etc. I had all sorts of ill-considered ideas at the time and no real guidance. It wasn't a _tremendous_ amount of code(running a simple count again, 31,145 LOC with whitespace/comments and 22,372 without), and Leo documented what it all did, but the tool couldn't suggest why it was fundamentally rotten, it just added process on top. By the end I wasn't really using Leo, because it wasn't solving my problems.
Looking back on it now, I have a style that can more naturally accommodate a literary programming approach because I'm more likely to write a straight-line solution first. But I would not rely on an external tool again as I don't want the dependency.
Perhaps the above would be less useful for more traditional code. Could you describe in more detail how you used it?
I have a long catalog of improvements I'd like to see made to Leo (e.g., its XML file format is version control hostile - try resolving merge conflits on several thousand lines of deeply nested machine-targeted XML; or try sending a pull request slathered with "sentinels"). Building such tools being out of scope for my day job, my re-imagined version won't be available any time soon.