> Cross platform (mostly). Linux (main target), Windows (experimental), Mac.
So, it is not targeting _terminals_ (like traditional libraries do, including ncurses), it is targeting entire operating systems. That suggests it might be relying on things it shouldn't be relying on, and possibly making a bunch of simplifying assumptions. Given that it also says:
Even terminal programs that just write to stdout depend on the OS to some extent.
On Windows, writing to a terminal window has different effects depending on the type of terminal. The classic Windows Console doesn't understand VT codes but there are API calls to do the same things (position cursor, change attributes etc). Recent versions of Windows Console can be told to understand VT codes, but require an API call to enable VT interpretation, then they understand a subset. Windows Terminal isn't the same as Windows Console. Then there's running things under Mintty, or Cygwin Terminal. Control+C behaves differently under those than under the Windows Console. Then there's running things under "winpty" in a Cygwin Terminal, which uses a hidden Windows Console and translates the resulting character matrix as best it can to terminal codes. There are other terminal emulators too.
Linux terminals have a lot of diversity regarding which VT codes they interpret, but at least there's a common subset for most of them. There's also some variation about codes for keypress, such as modifiers with function keys.
If it uses terminfo or termcap to work with various terminals, there are some OS dependencies regarding what capabilities are likely to be indicated.
The color capability varies. Sending 24-bit color to some terminals will not have the desired effect on others.
Unicode support varies among terminals, and this toolkit uses plenty of Unicode in the demos: Radio buttons, check boxes, box drawing characters, and the bar graph demo. What counts as a wide character (two cells) also varies.
Author of FTXUI here. Thanks for this answer! This is exactly the right reasons.
Every terminal compatible with the main subset of the various VT specifications should be compatible with FTXUI. If someone can prove this wrong, I must fix this ;-)
I own a Linux machine so this is the "main" target. Support is quite strong.
However, I don't own a Windows machine and it's quite painful for me to test on it. Windows support has been added by contributors mostly. That's the reason I say it is still "experimental". The various Windows terminal requires some Windows specific calls to enable VT interpretation, as explained above.
FTXUI can be compiled toward WebAssembly and executed against xterm.js terminal. This proves you can use FTXUI against a compliant terminal and not really depends on OS calls.
https://arthursonzogni.com/FTXUI/examples/
First, I suggest you try to use terminfo to properly ensure your terminal supports the "main subset" you've mentioned.
Second, I suggest you get in contact with Thomas E. Dickey, the maintainer of libncurses and terminfo, to see how he tests that library - on something like 15 different platforms.
Finally, over time, consider relaxing your set of terminal requirements - either with workarounds or with limitation of functionality for less-capable terminals.
I guess that working with more primitive terminals would make easier to use to drive UIs on embedded targets. The "terminal" Ftxui sees becomes just some LCD-painting primitives.
> That suggests it might be relying on things it shouldn't be relying on,
why ? If I have a choice between using a software that works on mac / win / linux, and one that works on those three + a ton of obscure architectures / old systems I'd definitely choose the first one - every additional platform supported invariably means that :
- you have to give up some features or reimplement them yourself (because you want that feature anyways) thus more possibilities of bugs, e.g. the amount of times I've seen people reimplementing badly some POSIX functions or stuff like clock_gettime on windows
- you have to duplicate code, say DLL / so loading if you have some plugin system (plus a mess of #ifdefs because of shit like macOS having a non-functional pthread_cancel, etc...)
- in some cases this may mean that the program author needs to add some run-time indirections, e.g. to call platform functions, which may have a performance cost.
- most likely the build system will be some terrible autotools mess which has not been updated in 15 years and does need 50 changes before working on a modern linux distro (had that merely a week ago) + a batch script for windows
- maybe some performance / technical improvements / simplifications would not be available because they would break compatibility with whatever EBCDIC IBM platform with two users, thus making the software worse, condemning us to the "state of the art" of the oldest supported platform
> If I have a choice between using a software that works on mac / win / linux, and one that works on those three + a ton of obscure architectures / old systems I'd definitely choose the first one
The first option probably means they are assuming the terminal behavior of the common terminal apps desktop users on one of the two/three main platforms are using. But if they're making that assumption, they might as well assume they can use a proper GUI and go with that.
The whole point of terminal programming is that you live within the abstraction which is the terminal. The fact that you're thinking about _systems_ rather than _terminals_ is the problem.
It's perfectly ok to have your app refuse to run on a terminal which lacks certain capabilities (annoying, but ok): You check for whatever you need from the terminal using Terminfo(https://en.wikipedia.org/wiki/Terminfo), and die gracefully if necessary. What's not ok is saying "Oh, I'm running on Windows, so XYZ must be true about the terminal I'm in, I'll just go ahead and assume that's the case".
Also - what you consider "obscure and esoteric systems", for many other people is just "systems".
I didn't bother to comment on this, but these are my thoughts as well. From briefly looking through the source code, it seems like the only way this library cares about the capabilities of the terminal is just checking if '256' or 'truecolor' is in TERM variable, and that's it.
> The whole point of terminal programming is that you live within the abstraction which is the terminal
I really fundamentally disagree that there is a "whole point" to terminals (or any other kind of things we use when we use computers). They are just tools which are able to perform things and one should really not ascribe any kind of semantic layer to them, as this is always a very very brittle abstraction. Just look at things in terms of what they are able to do, not what they are called.
Yes, that's correct. The whole frame is printed. As said above, as long as you are careful about flickering, this is fine.
This runs very smoothly. So I would say "premature optimization is the root of all evil" for now.
Note that in case of many events sent in a row, we don't render a new frame after every events. We first empty the list of pending events and only render once when it is empty. This ensure there are no lags.
If there is a need, this can be added easily. We just need to ensure this doesn't regress performance. Sending a single chunk representing the whole frame can sometimes be cheaper than sending several diffs at various locations of the screen.
Also, FTXUI support not only rendering to the alternate buffer (fullscreen mode), but it can also render inline. In case of terminal window resize, printing the whole frame avoids garbages to remain, which is nice.
That's a reasonable thing to do nowadays; I wrote a terminal text editor on a Sharp Zaurus in 02003 that worked that way, and it was fine. ssh -C removes most of the bandwidth cost if you're doing this remotely. You do have to be a little careful about flicker, and probably on mobile devices (like the Zaurus!) it costs you battery life.
Can I just say that the Ftxui source code is beautifully laid out, and is much more readable than I am used to seeing in Free Software projects? This library has absolutely exemplary presentation. [Edit: I see it uses clang-format. More projects should use that.]
Overuse of shared_ptr (what we call "Java Disease") is probably forgiveable in a UI toolkit. Even protected data members, often frowned upon in O-O circles, are probably OK, here. A note to the author (and, you know, everybody): the "#ifndef xxxx/#define xxxx" two-step is no longer necessary or useful; "#pragma once" is supported everywhere now.
I am terrified by #pragma once due to issues with relative include paths sometimes being mixed with absolute ones in various situations (none of which I personally cause). I recommend reading this post on StackOverflow by a GCC/autoconf maintainer on why he believes it is actually impossible to make it work correctly. I imagine it is thereby not in the standard, despite being widely implemented, "for good reason".
> (Historical note: The only reason I didn't rip #pragma once and #import out of GCC when I had the authority to do so, ~12 years ago, was Apple's system headers relying on them. In retrospect, that shouldn't have stopped me.)
I find the argument unpersuasive for the same reasons commenters there did. The case for C++ is stronger because the set of current C++ compilers (which nowadays have integrated pre-processor) is much smaller than of C compilers that could be encountered in the wild.
Doing complicated things with header files and paths is inherently fragile, pragma or no pragma. So, don't do them. Absent fragile header file or include-path games, there is no problem.
In the foreseeable future, #include will begin to fade from C++ code, and with it #pragma once, include guards, and any sort of worries about them. (Then we might start to worry about module name collisions, instead.)
Great project. One of its user git-tui looks well comparing to another git-cli (tig).
Wish we have some nice cross-platform GUI with modern c++ works across linux/windows/macos too, the only choice used to be Qt. Nowadays imgui seems a good alternative.
gtkmm works across all those 3 platforms, and had modern C++ before Qt did (e.g. actually using templates and libstdc++ instead of preprocessors and custom "standard" classes).
That's how Emacs works, although it uses (your choice of a set of) GUI toolkits. Not sure if you mean one single toolkit that covers both and looks identical. I guess Emacs would still mostly qualify.
No you are right, but that’s irrelevant to my point. The point being that Flutter is an example of a “react-like” framework to write UI that renders to multiple backends.
not really, it's important to me that we don't flatten things into a canvas rendering system, the idea is to remap abstract UI structures onto various projections (terminal, html, else)
I see. We’re talking about different levels of abstraction.
There are two approaches there. One is to project abstract UI onto platform controls (React Native), the other is to use “drawing” (Flutter). Both have their pros and cons.
The main drawback of abstracting over controls is that each platform works very differently and the abstraction quickly starts leaking. Also some of the targets don’t have any controls like the terminal. All you can do in a terminal is draw text in rows and columns.
That is why most (not all) cross platform frameworks abstract on the drawing level. Gtk, Qt, Flutter, Druid, Dear ImGui and many more.
I agree that it's not well defined, but I'd guess that a good rule of thumb is "fd 0 isn't isatty, and the process has the info (from env vars or arguments) necessary to open graphically" with an additional switch for overriding in either direction.
Htop is an example of a good TUI but it doesn’t really look like a traditional “drawn” GUI.
I think that is the distinction that parent is making.
In a TUI environment (ignoring mouse support) traditional GUI elements and layouts often don’t make sense. Like one of the examples with all the radio buttons in small boxes.
The good TUI apps like htop, newsboat, mutt, vim/emacs, tmux all have mostly line-and-column oriented simple layout primitives.
You can use the TAB key and spacebar to navigate TUI widgets, there's no need for mouse support. TUI layouts can definitely be overdone, but they're also useful in many cases.
Htop is great. Lately I've been reaching for a way to get an overview of process clusters (like firefox or vscodium like to make), and I can see there's an open feature request in htop for that. Would be a nice thing.
top is ok, but I'm always confused about the key-combinations, and the view is cropped (command line arguments etc.), so a scrollbar or some on-hover popup seems needed here.
Notcurses' principal author aspires for it to provide a widget toolkit for TUIs[1], but much of the demo and much of the commit log to date is heavily focused on graphical capabilities and lower-level details.
It may be that one day it ships with a robust widget toolkit; or perhaps someone else will build another library on top of notcurses with that as the principal focus.
I'm also impressed with the effort put into FinalCut[2], but notcurses seems to me to be paving a more interesting path forward for TUI development.
You probably just want to use the Remote SSH extension in VS Code. That should work a lot better than X11 forwarding, and it should be a lot easier than rewriting the whole GUI to use the terminal.
To reinforce that, even those giants that are slowly adopting Rust, do have employees seated at ISO C++ table and are major contributors to its compilers and IDEs.
> Cross platform (mostly). Linux (main target), Windows (experimental), Mac.
So, it is not targeting _terminals_ (like traditional libraries do, including ncurses), it is targeting entire operating systems. That suggests it might be relying on things it shouldn't be relying on, and possibly making a bunch of simplifying assumptions. Given that it also says:
> No dependencies
I'm even more worried.