For practical reasons it isn't always text, but more importantly (the first bit is a subset of this problem): almost every program has to understand (100%) the output of every other program it has to interact with.

If output were typed our applications could ignore (reliably) the bits they don't care about.

If output were typed we could search by output-type (e.g. diff produces patch-like-output, so don't bother piping to "xargs", but allow "less" (any type of input) or "git apply" (patch-like-input) or "patch" (patch-like-input).

Types could add a huge amount of searchability to CLIs -- and for some reason this isn't appreciated even though we've seen it in a much more limited context such as OOP.

(I think my original inspiration for this line of thinking is Paul Chiusano, and he probably deserves most of the credit, honestly.)

EDIT: Actually, pondering this a little more: All programs should be forced somehow(!) to carry a "source location" for the source of their output. This is such an annoyance in the daily life of a programmer if you perform a chain of commands like "git diff ... | grep FIXME | grep ^+ | ..." just to get a grip on the FIXMEs you've introduced. The output doesn't show line numbers, because... well, for no really good reason, really. It's just that transformation as text meant that we sort of lost that bit.

While I mostly agree with your thesis that types could be great in the shell, I think this particular bit is incorrect in a way that undermines your argument (not, I think, fatally). If a program's output doesn't match up with another program's input, a bit of sed (and maybe some awk) will usually do the trick. And if we added types, that would still be the case - output from one command would have to be the right type to be input to another.

Btw, I didn't mean that types should be the arbiter of "a | b", just that it should guide the search process. Obviously, you should always be allowed to really choose "c" even if the system determined "b". But our UI could be so much efficient if the system chose "b" most of the time.

EDIT: Classic case, btw: "find ... -print0 | xargs -0 ...". Unless we have pervasive and dependent types, how is any formalism going to know that?

Well, it's been sufficient. That doesn't mean we can't improve on it. My point was simply that assigning types doesn't actually change things - it is still the case that every program needs to be able to interpret whatever is generated by every program you want to pipe into it, or you need some kind of translation.

'Classic case, btw: "find ... -print0 | xargs -0 ...". Unless we have pervasive and dependent types, how is any formalism going to know that?'

Dependent types aren't necessary for that. They might be necessary for `find ... $MIGHTBEPRINT0 | xargs $MIGHTBEDASH0`

Edited to add - or wait, you mean the resulting command built by xargs may have a type dependent on the arguments produced by find? That's... messy, yeah.

Yes! Thinking about interactive CLIs... yes.

EDIT: I should add: My main point here is that types would constrict the search space to be reasonable, rather than seeing that "1200 commands are available, do you want to show them all?" prompt. I really think we could be doing much better in this space.

    grep -n '^+.*FIXME'

Note: on a phone so -n may not be correct.

EDIT: ... and it's actually possibly(!) even worse in the web world where you often have a ridiculous number of levels to your "transformation". Personally, I have never got "bower (source) -> grunt -> ts/(scala.js+closure) -> babel -> es5" (with source maps) to ever work even semi-reliably.

(I know we have too many things in this pipeline, but it's what we have to deal with.)

Basically, however widespread our personal workflows are, there are vastly more workflows we havent't even thought of that are essential to someone else's livelihood, for whom automatic numbering (or whatever) would ruin everything.

As for source maps, yeah, that's got room for improvement.

Hah!

EDIT: I had a really complicated thing here, but decided to delete it... because I don't think anyone understands any of this "computers" thing properly, including me. Hope you understand :).

EDIT#2: Is there a "flowers for you" emoticon?

C# supports extension methods, which are basically static helper methods whose first parameter is the type to which you want to attach the method. For example, I could create a 'ToTitleCase' extension method on type 'string'. Then, given a string instance, I can simply do 'myString.ToTitleCase()'.

One step further would be to add support for multiple parameters. For example, instead of listing all functions that accept an 'int' as an argument, you could ask for all functions that accept both an 'int' and a 'string' as an argument. Given 3 and "test", the 'repeat' function could be suggested, resulting in "testtesttest".

To make things even better, one could provide (explicitly or not) the return type of the function they're looking for. For example, if you're assigning to a variable of type 'int', and all you have is a 'string', then you probably want to see a list of all functions that take a 'string' as an argument, and return an 'int'. In that case, you would see functions like "parseInt" or "getLength". It's kind of like playing domino, where you already know the type of the input and output, and only need to connect them together. Keep in mind that you can't always transform one type to the other in a single function call, which means the system would also need to recommend combinations of functions, such as "parseDateTime + getDayOfWeek" when you need to transform a 'string' to an 'int'. Perhaps the name of the arguments could even be used as some kind of sorting heuristics (a 'string' variable named "date" surely should be parsed to some kind of DateTime at some point).