On the same general line: given a shell, which could analyze frequency of commands & make intelligent guesses for non-executable / invalid commands ("do what I intended, not what I've typed), I would toss my bash so hard against /dev/null, and salt whatever remained of it's config files.
Machines have to become much more intelligent before I would trust them with "do what I intended" if there is any chance that I lose my data if the machine misinterprets my intention. Leaky do-what-I-intended abstractions are the worst.
Autocomplete and autocorrect are "guess what I intended" functions. The problem with shells is that the UX is crap: User actions are often executed immediately (review time between typing and executing is a fraction of a second), and most importantly there is no undo. This makes guessing a bit dangerous. However, the fault is with shells, not with guessing. If you never instanty execute anything without review then the issue nearly goes away (more like regular programming than a REPL).
If it was less "take the whole of what I typed, rephrase it into something, and silently execute that", and more "fuzzy-autocomplete each token as I type it, and interactively lint the syntax to balance parens et al., so that by the end every token is almost certainly valid—and I can still then proofread what's there before executing it" then it's not nearly as scary.
Ideally, it would bias strongly towards blocking destructive activities, and hinting but not executing inferred commands.
Perhaps with something like "really" (where usage is similar to sudo) that would both override and train the fuzzy guessing. Training times might be a problem, even with a good algorithm, but perhaps if you got enough people to donate bash histories.
