Of course, it still has to be bootstrapped using the conventional bash approaches, but it's a first step.
This is somewhat like the approach that DEC took with VAX/VMS in the late 1970's.
The command line and arguments/parameters were declared external to the executable and known to the the shell (in this case, DCL) via the SET COMMAND command. It had its own little data description language. The shell did the parsing user interaction (recall, completion, etc) and provided utility functions to the binary to call to obtain params. So while the completion was not technically in the binary, a standard api was used to obtain it and the definitions were created side by side.
Worked very nicely and was extensible by 3rd parties and could be used to incorporate "foreign" binaries as well.
But I do think many important/useful binaries will never provide completions, or provide sub-par completion that people will want to improve, so we can't ever count on completions from binaries.
Recently I've been thinking some about how the LSP (https://microsoft.github.io/language-server-protocol/) might be useful for a CLI. A readline implementation could use something like a LSP for syntax coloring and completion.
Imagine having a consistent editing in your favorite shell, repl, and editor.
I took a survey of what people use here: https://lobste.rs/s/z96uyr/survey_what_shell_completion_scri...
That produced a link to this old project which tried to do something like what you want: a completion system that works for multiple shells (in this case, by code generation from a DSL)
https://github.com/mbrubeck/compleat
It's easy for you to say "completion should be specified in the binary" without actually specifying HOW. That's not straightforward, although I agree it's a good goal.
-----
Right now I'm implementing a bash-compatible completion API, and I'm actually getting pretty far without too much code. git-completion.bash is 3000 lines of code, so it's actually easier to emulate bash (with all its warts) than to write an entirely new completion system AND THEN write 3000+ lines of git completion in that hypothetical system.
https://github.com/oilshell/oil/issues/175
FWIW git is also the biggest completion script in the zsh source tree by a factor of 2 -- ~7000 lines IIRC with the next largest being around ~3000 lines.
Eventually Oil should have something better than bash completion, but the first step is taking advantage of ~40K lines of existing code in the bash-completion [1] project. I also looked at emulating zsh but that would be a lot more work than emulating bash.
However, I recently learned that zsh also emulates bash, e.g. the complete/compgen/compopt builtins, along with some global variables. So that is already technically true.
Still, I'd like to define some kind of "nicer" protocol that binaries can implement to get completion.
But that also sort of exists: I learned that when you do "ls --<TAB>" in bash right now, the script actually dynamically greps "ls --help" for flags! This is in contrast to how ZSH works -- it has a bunch of canned completions, which presumably suffer from a version skew problem.
Someone out pointed out "npm completion" in this thread, which is new to me (try it; it prints a bash script).
git is also adding 'git --listcmds' to help with completion. So the logic is partially in the binary, and partially in the shell completion script.
I think I've seen other ad hoc mechanisms along those lines too. Ah yes I recall that there was XML output for the Google flag parser [1], which is meant to be used for completion, although I'm not sure how commonly it's used.
So basically there are a lot of different systems. Hoping that someone will come along and produce a universal solution is probably wishful thinking. It's another "boiling the ocean" problem.
I think what would be nice is to collect examples like 'git --listcmds' and 'npm completion' (previously mentioned), i.e. places where the binary itself helps out with completion.
In those cases, more of the logic is shell-independent. It's not 100% and will never be, but it's nice to share as much as possible.
Also, it would nice to see other ways that commands support multiple shells. For example, git has both bash and zsh completion in its tree.
Though I'm pretty sure the zsh developers wrote their own completion that is much richer than the bash-like one in upstream.
I believe the problem is that "completion" in zsh has a higher standard. You can make a lowest-common-denominator solution already. I think that is OK for now, but zsh users might disagree.
I think the other popular interactive shell is fish (based on my survey), so I'd be interested in learning more about it. However I wouldn't underestimate the fact that probably 90% of people use bash.
Feel free to e-mail me if you want to talk about it more (address in profile), or you can chat with me here: