my i3-emacs integration

"it'll be too slow," they told me
march 2026

tiling window managers are wonderful. ultra-flexible text editors are also wonderful. for a spell, i thought i'd found the ideal solution in exwm… and i think it would have been, save for the fact that i use ordinary, graphical windows as much, if not more than, text buffers, and sometimes those windows are from dodgy programs (e.g., steam) that have trouble with EXWM's fancy input methods.

but i still like emacs a lot. hell, it switches light and dark mode on my machine (still)! so, inspired by such posts as \(\sqrt{-1}\)'s, i set out to get a common set of keybindings between emacs and i3, along with some sane defaults around opening terminals, splitting windows, etc.

i first tried a script with xdotool and emacsclient, as in the above-referenced article, and that worked… but proved to be too slow: i saw lags of up to a second timing the script gave a latency of 30 to 100 ms from invocation to exit, which is still pretty slow but not a dealbreaker. i still don't know where the rest of the latency came from. between sending input to emacs and it actually registering. i don't know if this is because of my emacs version, other packages, emacsclient weirdness, whatever, but that wasn't going to cut it. plus, it seems wasteful to launch a whole shell-plus just to register a keypress, especially for some of the most commonly pressed key combinations i use. so i did the only rational thing: i patched i3.

my objective was: instead of unilaterally handling commands bound via i3's bindsym, add an option to check the currently focused window to see if it's emacs, and if it is, pass the keypress event through to it. note that this feature has been requested in the past, and the i3 maintainers have deemed it to be out of scope. i would make this a more fully-fledged patch if that were not the case. if emacs decides "no, i3 should actually handle this," it can use i3-msg to route the action back.

i succeeded in that, though it might not be the most elegant thing in the world. if you know about xcb and want to give me advice, please! send me an email at web@khz.ac.

relevant i3 code

i3 uses xcb_grab_key() with owner_events = 0 on the root x window to intercept keys. the relevant code in src/bindings.c looks like all unpatched code snippets refer to i3 4.25.1, if you want to follow along.

172struct Binding_Keycode *binding_keycode;
173TAILQ_FOREACH(binding_keycode, &(bind->keycodes_head), keycodes) {
174    const int keycode = binding_keycode->keycode;
175    const int mods = (binding_keycode->modifiers & 0xFFFF);
176    DLOG("Binding %p Grabbing keycode %d with mods %d\n", bind, keycode, mods);
177    xcb_grab_key(conn, 0, root, mods, keycode, XCB_GRAB_MODE_ASYNC,
178                 XCB_GRAB_MODE_ASYNC);
179}

this code isn't super relevant, except that i3 entirely steals its bindings from anyone else by intercepting on the root window. if you're thinking that setting owner_events = 1 to allow event passthrough so we don't have to re-emit… that would be great, but that appears to instruct x to pass the event through to only the root window. which is not what we want.

in i3's handle_event() in src/handlers.c, if it gets an xcb event, it sends it off to a specialized handler based on its type:

1481switch (type) {
1482case XCB_KEY_PRESS:
1483case XCB_KEY_RELEASE:
1484    handle_key_press((xcb_key_press_event_t *)event);
1485    break;
1486    // ...
1487}

handle_key_press() (src/key_press.c) looks like this — it receives a keypress event, looks up a binding based on that event, and, if it finds one, runs the associated command: yes, i do know one of the lines is too long. i opted to leave it that way, as that's how it is in the i3 source. i should note, though: i3 has really nice source code! i found it very readable and pleasant to work inside.

12/*
13 * There was a KeyPress or KeyRelease (both events have the same fields). We
14 * compare this key code with our bindings table and pass the bound action to
15 * parse_command().
16 *
17 */
18void handle_key_press(xcb_key_press_event_t *event) {
19    const bool key_release = (event->response_type == XCB_KEY_RELEASE);
20
21    last_timestamp = event->time;
22
23    DLOG("%s %d, state raw = 0x%x\n", (key_release ? "KeyRelease" : "KeyPress"), event->detail, event->state);
24
25    Binding *bind = get_binding_from_xcb_event((xcb_generic_event_t *)event);
26
27    /* if we couldn't find a binding, we are done */
28    if (bind == NULL) {
29        return;
30    }
31
32    CommandResult *result = run_binding(bind, NULL);
33    command_result_free(result);
34}

notably, this function receives the original xcb_key_press_event_t from xcb, which (after a bit of reading and experimentation) i realized you could just re-emit diretly via xcb_send_event(). unfortunately, the window receiving the event will still lose focus, as i3 is intercepting key events globally. i haven't fixed this; let me know if you know how.

this looks like a reasonable place to make a change!

the patch

Binding struct changes

i decided to modify Binding (include/data.h) with an extra field to indicate a class of window which should, for that binding, receive events directly:

/**
 * Holds a keybinding, consisting of a keycode combined with modifiers and the
 * command which is executed as soon as the key is pressed (see
 * src/config_parser.c)
 *
 */
struct Binding {
    // ...

    /** Window class to use for key passthrough. Currently an exact string match. */
    struct {
        char *class;
    } passthrough;
};

i also modified the binding initialization to set up passthrough, if provided: there is, of course, associated cleanup code, which i've omitted for brevity. look at the patch file (linked at the end) if you want to see it.

/*
 * Adds a binding from config parameters given as strings and returns a
 * pointer to the binding structure. Returns NULL if the input code could not
 * be parsed.
 *
 */
Binding *configure_binding(const char *bindtype, const char *modifiers, const char *input_code,
                           const char *release, const char *border, const char *whole_window,
                           const char *exclude_titlebar, const char *command, const char *modename,
                           bool pango_markup, const char *passthrough) {
    // ...

    // XXX: should change this to be configurable, but I only care about Emacs, so.
    if (passthrough) {
        new_binding->passthrough.class = sstrdup("Emacs");
    } else {
        new_binding->passthrough.class = NULL;
    }

    return new_binding;
}

handle_key_press() now has to look at that setting and decide whether to pass the key event through. if bind->passthrough.class is set for that binding, we get the currently focused window, check its class, and if that class matches, we re-send the key event to that focused window with interception disabled (else it would just go straight back to i3):

void handle_key_press(xcb_key_press_event_t *event) {
    // ...

    DLOG("PATCH: checking if we should pass keypress through\n");
    if (bind->passthrough.class) {
        xcb_generic_error_t *focus_error;
        xcb_get_input_focus_reply_t *input_focus = xcb_get_input_focus_reply(
            conn, xcb_get_input_focus(conn), &focus_error);

        if (focus_error != NULL) {
            DLOG("PATCH: could not get focused window");
            free(focus_error);
        } else {
            Con *con = con_by_window_id(input_focus->focus);
            const xcb_window_t focus = input_focus->focus;
            free(input_focus);

            const bool should_pass =
                con && con->window->class_class &&
                strcmp(con->window->class_class, bind->passthrough.class) == 0;
            if (should_pass) {
                DLOG("PATCH: forwarding keypress (%d %s %s @ %d %d)\n", focus,
                     con->name, con->window->class_class, event->event_x,
                     event->event_y);
                event->event = focus;
                xcb_send_event(conn, false, focus, XCB_EVENT_MASK_NO_EVENT,
                               (const char *)event);
                return;
            }
        }
    }

    DLOG("PATCH: handling keypress normally\n");
    CommandResult *result = run_binding(bind, NULL);
    command_result_free(result);
}

modifying the parser

i3 includes a parser generator, which reads what appears to be an i3-specific dsl. since i want to specify that only certain bindings should be passed through, i had to modify i3's parser as well. the dsl is pretty self-documenting, so don't be afraid of doing this yourself. alternately, if you want to bug me to do it, email me.

the parser configuration for bindsym / bindcode (parser-specs/config.spec), after modification, looks like this:

# bindsym/bindcode
state BINDING:
  # ...
  passthrough = '--passthrough'
      ->
  key = word
      -> BINDCOMMAND

state BINDCOMMAND:
  # ...
  passthrough = '--passthrough'
      ->
  command = string
      -> call cfg_binding(..., $passthrough, $command)

this section of the parser config defines two parser states: BINDING (parsing a bindsym command, but we haven't parsed a keysym yet) and BINDCOMMAND (the same, but after we've parsed the keysym). the right way to do this, should i have wanted to have syntax like --passthrough "Emacs", would be to move to a new parsing state upon encountering this flag and eating the next token as passthrough. perhaps someday. i3's parsing dsl appears to accumulate variables via variable = <stuff> and passing that variable to a call command as a char* — non-null if encountered and null if not. hence, if the flag --passthrough appears while parsing, $passthrough evaluates to the string "--passthrough" rather than NULL. then if (passthrough) { /* ... */ } gets evaluated in configure_binding(), and the rest is history.

the emacs side

now that key passthrough works, all we need is a bit of elisp and life is good. a lot of this is heavily pulled from the \(\sqrt{-1}\) post linked above. basically, i want to integrate two actions: window movement and opening terminals.

window movement

to start, we need a way for emacs to send messages back to i3 if we try to move beyond an existing window:

(defmacro nausicaa/i3-msg (&rest args)
  "Call i3-msg with ARGS."
  `(start-process "emacs-to-i3" nil "i3-msg" ,@args))

when either moving windows or moving between windows, emacs should attempt to select one of its own windows in the given direction. failing that, it should instruct i3 to do so:

(defun nausicaa/emacs-i3-windmove (dir)
  "Select window in DIR, if it exists; if not, i3-select it."
  (let ((other-window (nausicaa/find-other-window dir)))
    (if (or (null other-window) (window-minibuffer-p other-window))
        (nausicaa/i3-msg "focus" (symbol-name dir))
      (nausicaa/do-window-select dir))))

(defun nausicaa/emacs-i3-move-window (dir)
  "Do some stuff to move window in DIR.

I should check out `evil-move-window' at some point."
  (let ((other-window (windmove-find-other-window dir)))
    (cond
     ((and other-window (not (window-minibuffer-p other-window)))
      (window-swap-states (selected-window) other-window))
     (t (nausicaa/i3-msg "move" (symbol-name dir))))))

nausicaa/find-other-window is a function that really just invokes the appropriate windmove command. i wrote it because my existing windmove commands have advice around them (placed there by doom, i expect) that allows them to select popup windows and the minibuffer, which i wanted to reuse:

(defun nausicaa/find-other-window (&rest args)
  "Pass ARGS through to `windmove-find-other-window'.

Exists solely so I can reuse `+popup--ignore-window-parameters-a'."
  (apply #'windmove-find-other-window args))

(defun nausicaa/do-window-select (&rest args)
  "Pass ARGS through to `windmove-do-window-select'.

Exists solely so I can reuse `+popup--ignore-window-parameters-a'."
  (apply #'windmove-do-window-select args))

(advice-add 'nausicaa/find-other-window
            :around #'+popup--ignore-window-parameters-a)
(advice-add 'nausicaa/do-window-select
            :around #'+popup--ignore-window-parameters-a)

arguably the right way to do this is to add that advice to windmove-find-other-window, which i might do at some point.

terminals

                      i am always launching terminals —
sometimes
          fifty
       a
          day.

i typically use mistty as a terminal, since it has delightful integration with the rest of emacs, but it tends to choke on more difficult text rendering tasks, for which alacritty is better suited. at any given moment, in any given directory, i might want to launch either of them, so i wrote a few scripts to invoke either mistty or alacritty from either emacs or i3.

i3 is configured to launch both mistty and alacritty, depending on context, using two scripts:

# start a terminal
bindsym --passthrough $super+Return exec mistty-create
bindsym --passthrough $super+Control+Return exec alacritty-create

if those keys pass through to emacs, emacs either launches a mistty session or just shells out to the script:

(defun nausicaa/launch-alacritty ()
  (interactive)
  (async-start-process "alacritty-create" "bash" nil "-c" "exec alacritty-create"))

(map!
 "s-<return>" #'mistty-create
 "C-s-<return>" #'nausicaa/launch-alacritty)

mistty-create is a shell script that tells emacs to open a new frame with mistty in it:

pkgs.writeShellApplication {
  name = "mistty-create";
  text = ''
    ${config.programs.emacs.package}/bin/emacsclient -e "(progn (other-frame-prefix) (mistty-create))"
  '';
}

alacritty-create instructs the current alacritty process to create a new window in the current working directory:

pkgs.writeShellApplication {
  name = "alacritty-create";
  text = ''
    if ! ${pkgs.alacritty}/bin/alacritty msg create-window --working-directory "$PWD"; then
       env -u INSIDE_EMACS ${pkgs.alacritty}/bin/alacritty "$@" >/dev/null 2>&1 &
       disown %env
    fi
  '';
}

this script has the wonderful property that, if invoked inside emacs, you get an alacritty window in whatever project directory you're currently in, yielding roughly equally ergonomic behavior between mistty and alacritty. it's great.

results

i3 and emacs play really nicely together now. if you want the patch for i3, it's here. i'll eventually post my full configuration with keycodes, but the above should be enough to get something working.

and if you're like me, and want to do i3 development on nix, here's what i used for a shell.nix:

{ pkgs ? import <nixpkgs> { } }:

pkgs.mkShell {
  nativeBuildInputs = with pkgs; [
    pkg-config
    makeWrapper
    meson
    ninja
    installShellFiles
    perl
    asciidoc
    xmlto
    docbook_xml_dtd_45
    docbook_xsl
    findXMLCatalogs
  ];

  buildInputs = with pkgs.buildPackages; [
    libxcb
    libxcb-util
    libxcb-wm
    libxcb-keysyms
    libxkbcommon
    xcbutilxrm
    libstartup_notification
    libx11
    pcre2
    libev
    yajl
    xcb-util-cursor
    perl
    pango
    perlPackages.AnyEventI3
    perlPackages.X11XCB
    perlPackages.IPCRun
    perlPackages.ExtUtilsPkgConfig
    perlPackages.InlineC
  ];
}