Cleaner extension point for evil/normal-state roaming over an animated terminal (anchor + wheel scroll)?

[amiorin]

Summary

When using ghostel with an animated terminal program (something redrawing
at ~30fps) under evil, the viewport re-anchoring in ghostel--anchor-window
fights any attempt to move point off the live cursor or scroll into scrollback.
I've worked around it with two define-advice forms, but they feel like they're
papering over something ghostel could expose (or do) itself — hence this issue.

My setup: Ghostty → Zellij → Emacs -nw (terminal Emacs), with evil +
a small evil-ghostel layer that keeps the terminal cursor synced to point
across evil state changes.

The two seams

1. The redraw anchor fights normal-state motion

Each redraw, ghostel--redraw-now re-anchors any window following the live
viewport via ghostel--anchor-window, whose final set-window-point snaps
point back to the terminal cursor (ghostel--cursor-char-pos). On a static
terminal redraws are rare, so this is invisible. On an animated one it fires
~30 times/sec and yanks point back the instant evil normal-state motion
(hjkl) tries to move it off the live cursor.

My workaround skips the anchor while point is parked off the live cursor in a
motion-capable evil state, and lets auto-follow resume on return to insert or
back onto the cursor row. FORCE anchors (paste/yank) are always honored:

(define-advice ghostel--anchor-window
    (:around (orig &optional window force) evil-ghostel-roam)
  (if (and (bound-and-true-p evil-ghostel-mode)
           (not force)
           (memq evil-state '(normal visual operator motion))
           ghostel--cursor-char-pos
           (/= (point) ghostel--cursor-char-pos))
      nil
    (funcall orig window force)))

2. Wheel scroll into scrollback gets snapped back

In insert state the redraw anchor is still live, so a mouse-wheel scroll into
scrollback is immediately yanked back to the live cursor. Normal state is
exactly where the advice above suppresses the anchor, so I flip the buffer to
normal state on any wheel event over a ghostel buffer. ghostel redispatches
wheel events to mwheel-scroll when it scrolls the Emacs buffer (mouse tracking
off), so that's what I advise — only switching from insert/emacs (normal/visual
already roam, and I mustn't drop a visual selection):

(define-advice mwheel-scroll
    (:before (event &rest _) evil-ghostel-wheel-normal)
  (let ((buf (window-buffer (posn-window (event-start event)))))
    (when (buffer-live-p buf)
      (with-current-buffer buf
        (when (and (bound-and-true-p evil-ghostel-mode)
                   (memq evil-state '(insert emacs)))
          (evil-normal-state))))))

Questions

1. Is there a cleaner, blessed extension point? Advising
   ghostel--anchor-window and mwheel-scroll (both internal) is brittle — a
   hook, an "anchor predicate" variable, or a ghostel-follow-mode-style toggle
   that callers could flip while point is parked off the cursor would let this
   live without advice.

2. Could roaming-off-the-cursor / scroll-into-scrollback become the default
   (or a built-in option)? The core friction is that the redraw anchor has no
   notion of "the user has intentionally moved point away." Is making that a
   first-class state something ghostel would consider, or is the
   always-follow-the-live-cursor behavior intentional for performance/simplicity
   reasons I'm missing?

Happy to turn any of this into a PR if you can point me at the shape you'd want.

[amiorin]

I have created the issue with Claude Code.

Long story short: I'm trying to recreate the neovim Terminal experience in Emacs with Ghostel.

[emil-e]

@dakra is more into the evil-mode side of things, but the current behavior is generally that we don't do the snapping in the emacs mode. However, evil-mode is a bit of a different beast, so perhaps we should generalize the behavior so instead of checking emacs mode, we check whether some other variable is set, i.e. ghostel-sync-window-state or something. emacs mode would set this to nil and so would whatever evil-mode hooks also. Then ghostel--anchor-window noops of ghostel-sync-window-state is nil or something like that. Just brainstorming.

For non-evil users, this is not really a problem today since point anywhere automatically enters copy (where terminal doesn't update) or emacs mode (which is guarded as per the description above.

@dakra How do emacs and copy mode interact with Evil? Does Evil replace them or coexist?

[dakra]

@amiorin sorry for the late response. Can you try PR #484 ?

[...] Then ghostel--anchor-window noops of ghostel-sync-window-state is nil or something like that

I made it similar to the redraw case and added ghostel-inhibit-anchor-functions so external integrations can control that better.

How do emacs and copy mode interact with Evil? Does Evil replace them or coexist?

they coexist. ghostel-evil sits on top.

[amiorin]

#484 works for me.