I would strongly argue that this thinking is putting the cart before the horse.
I don't use tmux, nor do I want to (though occasionally I have to use screen as a hack to keep programs running on remote servers, and I hate every second of it). Solutions like tmux arguably exist because terminals have poor UIs, and the terminal protocol is too weak to form the foundation for the kind of interactivity and statefulness provided by modern graphical UIs. If terminals were as powerful as, say, web browsers (not that I'm suggesting that anyone conflate them), the world would be a different, happier place.
I think Hyper  is going down the wrong path, but I strongly believe a new "terminal-oriented UI model/protocol" could be invented that would scratch every possible itch — good for text, mouse support, custom UI widgets, seamless remote connections, multiple screen regions — without sacrificing functionality at all.
There's been a lot more pushback on the scrolling decision than I had anticipated. It's not something I want in my terminal, but it seems that a simple feature like this is essential for others. Perhaps I should reconsider.
I worry that a "simple" feature like this may be overly complex internally. Performance with large amounts of output is also a concern. At least if we were to add support, it could be designed as a build feature and be removed completely if it were undesired.
Incidentally, AmigaOS' terminal design is worth exploring - it is a fascinating example of layering. And the AROS re-implementation, while not very clean, is also partially object oriented (in C; disclaimer: I wrote part of it) in how it layers the console units from the simplest to most complete (cut and paste support + scrollback). Even back in the 80's performance was good enough for this that AmigaOS started using dispatch-function based OO all over the place, and the AROS re-implementation of the console code uses that method, which is not at all the fastest way of doing it, but it's fast enough even on real M68k's.
Any modern PC is going to be at a minimum several hundred times faster.
All you need is to maintain a linked list of lines, and add code to free or reuse when you reach the maximum size of the scrollback buffer. If you want you can easily also just use a ring-buffer and wrap around if you want to set a size limit in bytes instead of lines, and just maintain indexes into it. It's trivial to do this in ways that doesn't cause performance issues.
A GUI terminal emulator for such an architecture needs at least two open file descriptors, one for the connection to the X server and one for the master side of the pseudo-terminal. The other 14, it turned out, were being used by VTE's scrollback buffer mechanism, which involves writing data that have scrolled off the top of the screen out to (temporary) files.
They had managed to reduce this, by rearranging the structures of the scrollback files, to 8 open file descriptors per emulation by 2011, and reportedly it will be soon down to 4.
Interesting tidbit #1:
It was mentioned elsewhere in this discussion that the alternate screen on many terminal emulators has no scrollback. This is because the programs that switch between primary and alternate screens aren't actually doing that as far as they are concerned. They are switching between scrolling mode and cursor addressing mode (see http://superuser.com/a/715563/38062 for details), the latter not really having the concept of a negative row coördinate.
The VTE widget was using twice the number of open file descriptors, because both scrolling and cursor addressing modes had scrollback files.
Interesting tidbit #2:
LXTerminal has this single centralized emulator process architecture, too. It has a rather nasty open file descriptor leak with which one can render LXTerminal completely unusable in about 1 minute (if one has an open file descriptor limit of 1024).
I don't need tabs (I use i3wm), I don't need splits or session management (when I need that I use tmux). But I do often need to scroll back on a temporary session when I didn't plan for it in advance. I open a lot of terminals. They're never all going to be tmuxed.
Frankly, removing two very important features for a lot of people (me included) and then claiming highest performance feels a bit like cheating.
This is the problem of Unix philosophy, because it applies differently for different people.
For example, I use i3, that is a tilling window manager, so I don't need tabs or split management for tmux. Actually, I removed tmux since I started to use i3 because using i3 features feels much more natural, since they apply to every window. However, there is only one thing that I can't have with i3 that is scrollback buffer, so while I don't need a terminal with tabs, I need support for scrollback buffer. Adding tmux just to get scrollback goes against Unix philosophy.
I agree. Rather than adding scrollback support to allacritty, maybe someone could write an independent program for scrollback support (a la dtach/abduco for detaching/reattaching)?
Such a program would be useful for all terminals which lack scrollback (alacritty, st, possibly others).
For the record I use st with dtach and dvtm; scrollback is supplied by dvtm, but it would be nice to decouple it some more.
- Device drivers feeds raw input to input-handler
- console.device manages a single console window. it receives raw input from the windows message port (courtesy of the input-handler) and "cooks" it into escaped character streams (which can include things like mouse reporting), and processes simple output that it translates into window output.
- console-handler receives the escape codes and interprets more complex sequences before passing the result on to the application that has the console open, and writes output back to the console.device.
Most of this would run in separate threads.
This lets any application open special filenames like "CON:" to open a console window.
Within console-handler, multiple different "units" are layered - in AROS (AmigaOS compatible open source), the basic (no copy and paste, no reflow, no scrollback) unit, the unit with copy and past and reflow, and the unit with above and scrollback, are layered on top of each other via inheritance, using a system-wide OO system modelld after Commodores old BOOPSI (basically a simple vtable based approach with a "catch-all" method dispatch entrypoint for user-added methods; it's not fast with deep inheritance hierarchies, but it's fast enough for this kind of use).
The copy and paste itself is implemented via a separate tool - ConClip - started on boot (and optional; if you don't start it you simply don't have cut and paste), which receives messages about what to cut and paste and writes it to the "clipboard.device", which by default writes each cut/copied piece of data to the CLIPS: volume as IFF formatted files, which by default maps to a directory in a ram-disk, but which can be re-mapped elsehwere. This all happens asynchronously, to accommodate cases where people e.g. remapped CLIPS: to a floppy and had to swap it in (rare, but possible if you had to deal with low memory situations).
This is something that frustrates me to date with Linux etc. - AmigaOS was far more decoupled, with clear, well-documented boundaries for people to hack on (e.g. several people wrote alternative console-handlers and console.device's that you could totally replace the original with to the point that any application that used terminal windows could be made to use your preferred console device. Even third party components tended to follow this approach (e.g. compression in AmigaOS is usually done via the XPK suite of libraries, which provided a third party API for opening compressed data streams, that let you plug in any compression algorithm as a library - as a result most apps in an Amiga system that supports compression can support most compression algorithms you drop in a system-wide library for).
I'm aware of BOOPSI, and the datatypes system which sounds similar to what you describe.
One problem on AmigaOS was(/is?) the lack of packaging and dependencies, e.g. installing many programs on a fresh copy of Workbench won't work, due to missing libraries, etc. Thankfully that's easier to manage these days by scouring Google and Aminet, but it's still manual.
Interestingly, I've found Amigas to become more stable over time, unlike e.g. Windows where some people recommend formatting every year or so to remove cruft. The more stuff you install in Workbench, the more libraries, etc. you accumulate, so the fewer problems you encounter trying to install/use other things. I'm not sure if this is a consequence of the OS design, or from developers bending over backwards to avoid problems (e.g. conflicting names, etc.)
> One problem on AmigaOS was(/is?)
Is, sort-of. Package managers didn't enter the scene until much later, but Commodore did release Installer, which while not a package manager provides a s-expression based mechanism for describing installation flow.
It's alleviated because Amiga libraries tends to very strictly insist on backwards compatibility, so you should generally be able to drop a newer version of a library over an older version and things will keep working (and the libraries and all compliant binaries contains version numbers).
But of course the community today is very small, and was smallish originally too, and so it's gotten easier and easier to deal with.
If there was to be a resurgence (there is new hardware but it's expensive niche PPC hardware; AROS runs on pretty much "anything", but is incomplete), it'd need a lot of big overhauls - in particular memory protection (some work is ongoing but it's hard due to AmigaOS APIs relying a lot on pointer passing) and SMP, but also lots of tooling we take for granted today like package management.
I'm not holding my breath for that, but I do wish more AmigaOS ideas will get picked up elsewhere. Linux still feels like a hodge-podge in comparison.
There are also simpler tools which can provide scrollback support.
The whole Unix tty subsystem goes against the "Unix philosophy", that is why it was completely scraped in Plan 9.
That principle is often misapplied, and I think that's true here, too.
The "do one thing" about Unix is really about composability (e.g. "find" doesn't need to sort because you do "find | sort"), but you don't compose a terminal app with anything.
A terminal app that has terminal features doesn't violate any principles of simplicity.
It's rarely a significant problem in practice, but it annoys me in principle!
If you do "blah | sort", then "sort" could ask its upstream processing node whether it supported sorting on the requisite fields, and "push down" the necessary sort-order descriptor into the "blah" step.
That requires two things: That the pipe API sets up a communications channel between the two programs in a way that makes them aware of each other and able to exchange information; and secondly, that the pipe protocol is based on typed, structured data. I want both things.
Imagine if you had that, then you could conceivably also do:
psql -c "select firstname, lastname from foo" |
sort -f lastname
select firstname, lastname from foo order by lastname
The inability to do this sort of thing really a product of a failure to modernize the 1970s text-oriented pipe data model. I believe PowerShell (which I've never used, only read about) provides a mechanism to accomplish this sort of thing, at the expense of being extremely Microsoft-flavoured.
I don't think there's anything even vaguely scifi about those abilities, but the Unix world is hampered by a curious reticence to innovate certain core technologies such as, well, Unix itself. That's why we still have tmux and such.
Everything's highly composable, e.g. I can switch out bash for zsh, fish, etc. I can switch out dtach for abduco. I can switch out dvtm for tmux or screen. I can switch out st for xterm or urxvt. And so on.
Adding an extra layer for scrollback, separate from a multiplexer, wouldn't disrupt anything, and would provide more flexibility for composition.
I shaved off my neckbeard, I'll have you know! ;)
My setup's no more outlandish than using tmux or screen, except instead of typing `tmux` or `screen` I type `shell`, which aliases a `dtach dvtm` one-liner (with a few options sprinkled around, so I don't have to bother with config files).
The point is that none of these applications care if/how they're composed; if I want to add in or swap out something, it's just a change that one-liner.
Not so if, say, my terminal application were hard-coded to rely on tmux, as some sibling comments have suggested.
In AmigaOS it was fairly common for applications to reuse the standard console handler - the same one used for the shell - to provide a custom console or even editor windows etc.. For minimal integration all it takes is to read from/write to a filehandle.
That said, though as I noted elsewhere, even AmigaOS got scrollback in the console handler by '87 - the extra code measures a few KB; it'd be more hassle than it was worth to split it out.
It still depends how these features are implemented.
For example in dvtm scroll back history is made searchable by piping it to $PAGER. Similarly, copy mode is implemented by using $EDITOR as an interactive filter.
There are some cases where tmux could help me, but my X session and terminal emulators are stable enough where I'm not worried about them crashing and interrupting my shell session. As such, I have no need for tmux, and using it just for terminal emulator scrollback seems hamfisted.
The only use case I can think of is running large distribution updates which might potentially pull the rug out from under your graphical session, but I tend to run those on a non-X terminal if they look sketchy.
Edit: I'd totally consider using tmux as an alternative to X on my Raspberry Pi, but if I have X/Wayland, It doesn't offer me much.
As much as I like tmux (although I exclusively use it via byobu), the single most annoying thing is that it won't let different viewers see different content. (eg start two terminal emulators, run tmux in both and any switching you do in one affects the other. There are supposedly elaborate workarounds, but far too much effort.)
Ctrl-ALT-F2 would take a miracle to work. Shift does as you describe, but the net effect is still nowhere what is intended.
Also note I am the same user. The functionality I want is how byobu behaves with screen. eg you can start 3 xterms, in each one run byobu. And each one can jump around as they see fit all sharing the same screen session. No work, no fuss and exactly sane.
The tmux behaviour baffles me. I can't understand why anyone would want all their viewers to change in sync. Short of a classroom demo environment, it really doesn't make sense as a default.
I'd be delighted if by default byobu did whatever was needed to make tmux behave usefully.
Am I the only one who runs more than one terminal at a time? What happens now is I start an xterm on monitor #1 and start byobu within that. Various windows (whatever you switch amongst on pressing F3 and F4) are started - eg one might be client code running, one might be server code, one might be a database server etc. But sometimes for example I want to look at the client code output and server output simultaneously. At that point I switch byobu to the client window, and start a new terminal on monitor #2 and tail logs or whatever is appropriate. It is an annoying pain that I can't just run byobu and switch as I see fit.
In any event this is a multi-year frustration for me. People keep coming with convoluted workarounds (pointing tmux to tmux as far as I can tell) which it then isn't possible to figure out how to apply to byobu. All the while I wonder why two xterms running next to each other would ever want to be in lockstep by default?
for i in 1 2 3 ; do xterm -e byobu & ; done
tmux new-session -t main -s main2
The problem isn't that it is possible in theory. It is annoying in practise, unless you only have one machine and only need to do all this once in one system.
Even if tmux is in theory a better solution, it's not such a radically better solution that the benefits outweigh the switching cost. Terminals with scrollback have been around at least 30 years at this point , and scrollback is by far the majority use mode. You're asking 95%+ of your potential userbase to spend a bunch of time retraining themselves for little or no benefit.
You're also up against a variant of Jakob's law:
Scrolling down and back is a key part of the web user experience. Even if every terminal emulator got rid of scrollback, most of your users would still have scrolling back baked into their brains. When they want to see something that just scrolled past, they are going to do what they do in a web browser, which is to look for a scroll bar, hit a scrolling key, or perform a scrolling gesture.
That's not to say you shouldn't try radical things. But if you want user adoption, you have to make sure the benefits you offer are very much larger than the costs you impose. So radically different UI can't be about as good as the existing one; it has to be radically better.
 And of course terminals are made to emulate teletype machines, which had infinite scrollback to begin with.
I generally run five or six desktop terminals, only one of which is running screen locally. I more normally run something like screen to get persistence across suspends on an SSH connection. I usually don't do much side by side stuff in terminal so tmux isn't a big win for me. Also tmux didn't run on cygwin for the longest time, and I expect the same experience on all my platforms.
However the big tmux users in my life would still use it as is.
I am using firefox with pentadactyl, and the firefox is configured as a minimal terminal like st. Basically, firefox becomes my GUI terminal. But what I want is a much lighter weight GUI terminal than firefox.
I am using three terminals, st for regular use. mlterm for its image support (sixel). and firefox for heavy-GUI needs. Alacrity seems perfect to replace st now.
Thanks for good work.
Just do it. I doubt it will be hard to beat tmux performance, in my experience it's slow like molasses and I only use it if I must - remote persistency, screen sharing, etc.
I have the opposite impression. I actually anjoy time spent in my tmux + vim setup while I hate the time spent in a web browser. Web browsers seem bloated to me and I always feel like I'm forced to use the mouse too much while using them. I've heard about vimperator, but every time I've tried it seemed poorly integrated (the authors did an incredible job nonetheless).
The only drawback of delegating scrolling to tmux that I see is that my experience running tmux locally and using ssh to use a remote machine with another (remote) tmux session wasn't great. Keyboard shortcuts were usually caught by the local machine and not by the remote one. So I just don't run tmux locally when I want to ssh into a remote machine. Problem solved more or less.
Whilst I totally agree with you, I think Tmux is a lot like vim in its power. Along the same note, I'd wager a gTmux, much like gVim would be q real nice way to multiplex terminals when we get the UI to beat the TUI.
If you want keyboard only, I don't think GUIs can beat TUIs. But yeah, there's a learning curve.
The problem with GUIs is a very display-session centric view. tmux/vim work fine over SSH, Terminator/gVim don't. With tmux, your sessions are separate from your terminal instance. If your X session crashes, depending on how you started tmux, you just have to relaunch a terminal and reconnect to tmux. This is pretty invaluable. So this separation is powerful, and IMO very Unix-y.
For the thing to replace the TUI, I'd expect something in between current terminals and X11. With the simple, limited (and thus easily remotable) data of the current terminal emulators, but with much more drawing capabilites than the current grid-limited ascii art.
I came to this idea when trying to get vim up to a full IDE. Trying to get even half of netbeans' interface into vim just takes so much space in the ascii grid. And anything dynamic moves half the screen a shitton.
Although you could ditch it and use MOSH instead.
To me, it's a deficiency in the lack of session management for SSH. All that SSH gives me is a two-way pipe to the other server's I/O. That's simple and elegant, but why does it create a new pipe every time? It's connection-oriented, which is a concept that hasn't seen any innovation since the 1970s.
My preferred innovation here would be a local shell that had remote access. Rather than pipe I/O to the remote shell, give me a local shell which happens to execute its verbs on the remote machine, and let the remote file system simply be a volume. All my session state (including command history) can be local, there's no need to keep that on the remote host. A remote host is just another context.
I believe Plan 9 tried something similar, but very few people have picked up on its innovations.
~ $ cd /ssh:tol-eressea:.
/ssh:tol-eressea:/home/db48x $ ls
db48x.net libvirt-sandbox rpmbuild zone.sh zone.sh~ zone.txt zone.txt~
MOSH is also nice; it decouples the program you're running from your connection, just like screen does, and it also uses udp instead of tcp so that it doesn't have to worry about dropped connections. Even your client's ip address can change and everything keeps working.