Richer command line interfaces

21 Jul 2016

This post is preceded by a description of two types of command line interfaces. The more command line-like of these two types integrates into a normal shell session (python, telnet, bash) instead of using the alternate screen to present a fullscreen text user interface (top, vim, emacs). Sometimes programs blur this boundary with fancy formatting in a line-oriented command line interface, carefully avoiding overwriting previous user inputs and command outputs. On-keystroke autocompletion, multiline editing, inline documentation and syntax highlighting can be implemented with these techniques, making command line interfaces more useful.

But the terminal can be pushed still further! Here are three features that I’d like to see used in more programs (or I’d like to better understand their tradeoffs).

Checking vertical space before scrolling

When writing this kind of program it’s useful to consider two separate canvases available for drawing the current line and interface widgets like autocomplete suggestions: space that can be drawn in without causing the terminal to scroll down, and space that can be drawn in that would cause the terminal to scroll, but would keep the cursor onscreen.

Applications

Most autocompletion doesn’t seem to measure and use only the number of usable rows below the cursor, instead providing the same number of completion suggestions regardless of the amount of space left in the terminal.

Fish seems to force 5 rows by default, which may push terminal content up or may not be enough to fill the window:

while Python Prompt Toolkit does a nice job, still forcing 5 rows:

bpython measured the visible space but used too large a minimum completion box size until a few weeks ago:

Implementation

The larger of the two areas available for completion suggestions in the diagram above is as tall as the height of the terminal, which in a unix-y environment is a single cryptic TIOCGWINSZ ioctl() call away. But figuring out how much space is available without scrolling depends on where in the terminal the prompt has been printed. Finding this is a bit tricky. We can get at it by querying the terminal for the cursor’s position with \x1b[6n, but it’s a bit of a hassle because the response comes back from the terminal on stdin and needs to be distinguished from user input.

$ echo -ne "\x1b[6n"; read -sd R POS
$ printf "%q\n" $POS
$'\E[14;1'  # cursor is in row 14, column 1

Once we know where the cursor is, we can do some row/column arithmetic to find the region that can be used for completion without causing a scroll. Keeping previous lines of history visible is generally useful, and particularly so when livecoding for a group that is trying to follow along.

Handling terminal resizes

Many terminal emulators allow you to resize a session already in progress – something that would have made no sense in a traditional video terminal. Each time the user changes the terminal size a SIGWINCH signal is sent to the program and the terminal emulator moves the terminal contents around. The signals may be debounced a bit as they pass through additional layers of terminal emulation like tmux and GNU Screen.

Fullscreen, alternate screen-using programs don’t care how this content gets moved around because they can completely rerender after the resize. But a line-oriented command line interface should care more: if the resizing behavior can be predicted, editing artifacts can be prevented from polluting terminal history.

Applications

Here’s an example of a pathological case: a wrapped cursor causes the multiline input to be displayed a row too low on each resize:

The desired behavior here would be to not fill the screen with repeated text, but this is surprisingly difficult to achieve. Tools I’ve worked on suffer from similar problems. I’ve worked hard to minimize them, but I haven’t been able to get it anywhere near perfect, particularly for cases like “the user quickly narrows, widens, then narrows the terminal.” Single resizes of any kind aren’t so bad, and the result of any number of consecutive height changes is predictable.

Resize behaviors have both dramatic (history rewrapping) and subtle (which lines are shown or hidden when changing height) differences in different terminal emulators, so querying the cursor position after resize is useful for checking to see how content has been moved around.

Implementation

I spent a considerable amount of time trying to get this right in bpython, but it still fails in many cases. Later implementations for other projects look more promising but are less complete. My first suggestion is to debounce window size changes: wait to redraw until no new SIGWINCH signals come in for a bit. This sacrifices responsiveness in order to mess up history less: if you’re wrong, at least you’re only wrong once, ruining a few rows instead a few dozen.

Changing the height of a terminal window moves text in a predictable way, with a few edge cases that can be detected by checking the cursor position after the move.

An example of different behavior requiring a cursor check to detect:

Changing window width on terminal emulators that wrap text is trickier, particularly while text from a previous program remains visible, since the length of these lines can’t be known. But again, checking for the new cursor position should generally give enough information.

To get an implementation of this correct, it would be helpful to test multiple popular terminals for wrapping behavior. I briefly looked into orchestrating iTerm for this but found tmux easier to script and more portable. For that project I use tmux to check that ASCII diagrams correctly describe resizing behavior:

def test_wrapped_undo_after_narrow(self):
    self.assert_undo("""
                             +------+  +------+
    +------+  +-----------+  |$rw   |  |$rw   |
    +------+  +-----------+  +------+  +------+
    |$rw   |  |$rw        |  |>1    |  |>1    |
    |>1    |  |>1         |  |>stuff|  |>@    |
    |>stuff|  |>stuff     |  |abcdef+  ~      ~
    |abcdef+  |abcdefghijk+  |ghijkl+  ~      ~
    |ghijkl+  |lmnopq     |  |mnopq |  ~      ~
    |mnopq |  |>@         |  |>@    |  ~      ~
    |>@    ~  ~           ~  ~      ~  ~      ~
    +------+  +-----------+  +------+  +------+
    """)

It would be great if there were a really solid implementation of this approach, but it alone might not be worth the effort. Users of command line interfaces are trained to hit ctrl-l to request a clear or redraw as a workaround, and are already used to aggressive resizing messing up their history. However, the same machinery is useful for the next technique.

Rewriting previous lines

If we remember how many lines have been printed, we can know it’s ok to write to those lines. For example, autocompletion suggestions could be shown above the current line instead of below:

$ echo hello
hello
$ prompt_with_autocompletion_above
> s = "a"
> s += "b"
> s += "c"
> s
'abc'
> s.en|

$ echo hello
hello
$ prompt_with_autocompletion_above
> s = "a"
> s += "b"
> s+--------+
> s|encode  |
'ab|endswith|
> s.en      |
   +========+

To do this, it’s necessary for the program to model how much of the terminal can be written to and restored:

Careful attention to the terminal window size is even more important now that there are so many potentially wrapping lines.

Applications

In order to keep more history onscreen, it can be nice to show a completion box above instead of below text:

This is off by default in bpython because it’s surprising and a bit annoying if you wanted to see previous lines. But the more general feature of writing to previous lines enables all sorts of features like editing previous lines of a REPL session, reloading changes to a module, and undoing the last command:

Undo in this kind of stateful, side-effecting tool is pretty far out, a wackiness I hope to write about soon. But rewriting the text above the current line can also be useful for side effecting programs that can’t easily recreate previous states: it’s possible to clean up error output, collapse long command outputs, and annotate previous commands currently onscreen based on the current command. Alternate screen-style interfaces can be displayed in portions of the screen above the current line this way in addition to below.

Implementation

As required for dealing with window size changes, our command line program must be responsible responsible for simulating normal terminal behavior like wrapping in order to predict the contents of the terminal so they can be rerendered.

Each time I’ve implemented this I’ve written the terminal emulator lines wrapping calculations myself. After a too-naive approach the first time led to bugs, I’m most interested in an approach using a headless terminal emulator to make these decisions. I haven’t yet found one that implements linewrapping quite the way graphical terminal emulators do and would appreciate suggestions.

But even ignoring terminal window resizes completely sometimes produces useful results, in this simple Haskell REPL wrapper:

Of course there’s a limit to how much of history can be changed: the scrollback buffer cannot be modified! So bpython writes a warning offscreen that scrollback history is no longer consistent with the history that appears after it:

Call to action

If you know of other programs that use these techniques, I’d love to hear about them. If you know why they aren’t common, I’d love to hear that too! And if you try implementing them, I’d love to chat about implementation. It’d be great to see them in a tool like readline we can all just use, the way Python Prompt Toolkit makes adding so many fancy features to a REPL very simple.


Thanks to Amjith Ramanujam, author of the wonderful pgcli and other interactive REPLs, for feedback about these features and for pointing out that Python Prompt Toolkit already implements the first one.