Cursor shape support

[dabeibao]

The escape sequence is:
ESC[N q

N is between 0 and 6.

[achernya]

What is the context of this?

[dabeibao]

Please see the link:
https://gist.github.com/fnky/458719343aabd01cfb17a3a4f7296797?permalink_comment_id=5165100#gistcomment-5165100

Many modern terminals (including Windows Terminal, mintty, etc.) support changing the cursor shape dynamically — and this pairs beautifully with Emacs Evil mode, so you can visually distinguish between Normal and Insert modes (like in Vim).

Normal mode:

Insert mode:

[achernya]

Is there any more authoritative documentation of these escape sequences?

[dabeibao]

What about this: https://invisible-island.net/xterm/ctlseqs/ctlseqs.html

Search 'CSI Ps SP q' in the page:

CSI Ps SP q
          Set cursor style (DECSCUSR), VT520.
            Ps = 0  ⇒  blinking block.
            Ps = 1  ⇒  blinking block (default).
            Ps = 2  ⇒  steady block.
            Ps = 3  ⇒  blinking underline.
            Ps = 4  ⇒  steady underline.
            Ps = 5  ⇒  blinking bar, xterm.
            Ps = 6  ⇒  steady bar, xterm.

[achimnol]

We have long standing PRs and issues for this feature.

• Cursor Color and Style changes #352 (duplicate: Cursor style not handled correctly #1084)
• Apply Clang formatting to #1167 and resolve merge conflicts #1292
• Closed in favor of #1292 #1167

[eminence]

I'm looking at the impl for Dispather::getparam and it looks like doesn't support returning values less than 1. But 0 is a valid cursor shape.

[dabeibao]

Thanks. The defaultval should be 0.

[dabeibao]

* [Apply Clang formatting to #1167 and resolve merge conflicts #1292](https://github.com/mobile-shell/mosh/pull/1292)

Seems this commit already implements it. But the title is somewhat misleading...

[eminence]

Something's still not quite right when the terminal is reset. Try this:

printf '\x1b[6 q\x1bc'

[dabeibao]

Something's still not quite right when the terminal is reset. Try this:

printf '\x1b[6 q\x1bc'

Do you mean it also needs to reset the cursor shape after a terminal reset?
I tried Windows Terminal (without mosh), and seems it doesn’t reset the cursor shape either.

[eminence]

You're right that Windows Terminal doesn't reset the cursor shape after reset. But I tested xterm, gnome-terminal, alacritty, and ghostty, and they all reset cursor shape after reset. So I think we should too.

[dabeibao]

I have tried this in both mintty, xterm, gnome-terminal and alacritty (in WSL though).
These terminals reset the cursor shape after \x1bc, both with and without mosh.
But for window terminal, it never resets the shape even without mosh.

So it seems that this behavior depends on the terminal itself?

Which terminal do you see different behavior with mosh? I can have a try with it.

[achernya]

mosh claims to be xterm and therefore should act like xterm.

[eminence]

@dabeibao I'm not seeing that. Here's a more detailed screenshot showing my test (using commit 4ead896)

Running printf '\x1b[6 q' shows a steady bar cursor, as expected:

But then running printf '\x1bc or reset still shows the bar cursor:

But when I do the above steps in a non-mosh terminal, the cursor resets back to a blinking block cursor (the default)

[eminence]

The problem seems to be that the cursor shape is initialized to -1 when the terminal is initialized or reset. But a value of -1 will not be synchronized with the client, so after a reset, the server has shape=-1 but the client has shape=6

  if ( f.ds.cursor_shape != frame.last_frame.ds.cursor_shape && f.ds.cursor_shape != -1 ) {
    snprintf(tmp, sizeof(tmp), "\033[%d q", f.ds.cursor_shape);
    frame.append(tmp);
  }

I think I would just get rid of the init to -1 and init to 0 instead

[dabeibao]

The problem seems to be that the cursor shape is initialized to -1 when the terminal is initialized or reset. But a value of -1 will not be synchronized with the client, so after a reset, the server has shape=-1 but the client has shape=6

  if ( f.ds.cursor_shape != frame.last_frame.ds.cursor_shape && f.ds.cursor_shape != -1 ) {
    snprintf(tmp, sizeof(tmp), "\033[%d q", f.ds.cursor_shape);
    frame.append(tmp);
  }

I think I would just get rid of the init to -1 and init to 0 instead

Yes, thanks for finding this.
I now understand why I couldn’t reproduce it. I was using a private commit 6a533d6 (which includes other changes like cursor color and strike-through, so I didn’t push it to this PR yet).
It uses a counter initialized to 0, so it occasionally bypasses this bug. 😊

I tried to initialize cursor_shape to 0. But it still has some bug:
After reset:
Without mosh: steady block.
With the fix: blinking block.

In xterm, the default cursor_shape is 'steady block'. And without mosh, after reset it is still 'steady block'.
But if we initialize cursor_shape to 0, after reset mosh will detect the change and send a '\033[0 q', which changes the cursor shape to 'blink block'.

[dabeibao]

I tried to investigate further. Seems the problem is in the reset (\033c) itself.
Without mosh, when xterm receives the \033c seq, it calls VTReset then ReallyReset to reset cursor blink:

#if OPT_BLINK_CURS
    screen->cursor_blink_esc = 0;
    SetCursorBlink(xw, screen->cursor_blink_i);
    TRACE(("cursor_shape:%d blinks:%d\n",
	   screen->cursor_shape,
	   screen->cursor_blink));
#endif

screen->cursor_blink_esc = 0; will clear the blink state previously set by escape sequence.

However, it appears that mosh consumes the reset seq and doesn't forward it to the terminal.
After modifying the code to forward the reset seq and I did see the cursor reset to a steady block in xterm. However it also caused unexepcted behavior (e.g. the shell prompt being rendered as blank spaces).

Another issue with xterm (so as gnome-terminal) is, the initial cursor shape is configurable. User can set both the shape and blink state. Thus mosh cannot assume the initial cursor_shape value.

So from my perspective, we may have the approaches:

1. Set default cursor_shape to -1
   After reset we keep cursor shape. If users never change the cursor shape, they will never see any differences.

2. Set default cursor_shape to 2
   This matches xterm's default configuration. But it could override/break user customized settings after reset sequence.

I personally prefer option 1 as it will not impact the users who don't need this sequence.

[eminence]

Ok, I think I probably also lean towards approach 1. It does seem the safest and we can always tweak this later after more experience with this change.