Super + K lists Lua function bindings, but selecting one did nothing. For example, a user who zooms in, searches for Reset zoom, and presses Enter gets no reset or message; the script exits 1. Five of the 232 default bindings on Omarchy 4.0.2 are affected: Zoom in, Reset zoom, and the three universal clipboard actions. User-defined function binds have the same failure.

The menu replays the Lua config to recover each action, but a plain Lua function leaves no expression to replay. Hyprland reports these binds as __lua with a registry ref. When an expression is available from the source, the expression still wins. Otherwise, the menu records the bind's modifier mask, submap, key, keycode, and description. The record is stored base64-encoded, because later steps rewrite text in that field for display (key names for code:NN, the Omarchy bin path).

A config reload can assign the same registry ref to a different function. At dispatch time, the menu rereads hyprctl binds and requires exactly one __lua bind to match the recorded identity. The menu calls the ref currently attached to that bind through the existing hl.dispatch path, and refuses a blank or nonnumeric ref.

A reload also starts Hyprland's Lua state over. Before the lookup, each pick writes a mark of its own into a table in the Lua registry, and the call checks for that mark. A reload between the lookup and the call removes the mark, so the call raises an error and nothing runs. Overlapping picks keep separate marks, and a pick removes its own mark whether it is called or refused.

The picked row is now matched through the environment instead of awk -v, which reads a backslash in a description as an escape and then matches no row. This affected every bind kind. When two rows read the same and one of them is a function bind, the pick is refused, because the menu cannot tell which row was picked.

The cache version moves from v14 to v19 so cached rows with blank dispatcher fields are rebuilt. Open pull requests already use v15 to v18.

PR #11733 is a broader rewrite of the keybindings scan across 11 files. This PR keeps that work separate and fixes the current script.

Testing

• bash test/shell.d/keybindings-menu-test.sh: all 33 passed (16 existing and 17 new). The stubbed hyprctl dispatch runs the dispatched expression in lua, so the tests assert which function was called.
• Each new test fails against a version of the script with the matching defect: no mark check, one shared mark, no mark cleanup, a cleanup that empties the table, a lookup that ignores the submap or accepts two matches, awk -v for the picked row, the first of two identical rows dispatched, and a picked row that is not dispatched.
• bash test/shell.d/bin-style-test.sh: passed.
• git diff --check: passed.
• Live on Omarchy with Hyprland 0.56.2, running the script with only the picker stubbed: Reset zoom changed zoom_factor from 1.5 to 1, Zoom in changed it from 1 to 2, and two menus started together both exited 0. No mark was left in the Lua registry afterwards.
• ./test/shell: 279 of 290 test files passed. The 11 failures also fail on the unchanged base commit 821ae589; four of them require an omarchy-pkgs checkout that is not present on this machine.
• ./test/cli: one check fails, "vscode generated theme references current theme file". It fails the same way on 821ae589 (#13131).

Not included

• A reload between writing the mark and reading hyprctl binds refuses that pick once. Nothing runs, and picking again works.
• A reload while the menu is open keeps the row valid as long as one bind still has the same chord, submap, and description. The menu calls whatever function that bind has now, as the key press would.
• Function binds whose rows read the same are refused: the same chord and description in two submaps, or a press and a release bind. The picker returns only the row text. Telling such rows apart needs a row id carried through the picker, which is a separate change.
• A function that returns { ok = false } still makes the script exit 0, because hyprctl dispatch prints ok for it.
• The source cache is keyed by modifier mask, description, and key. A function bind that shares all three with a bind the source resolves takes that bind's expression. That keying predates this PR.
• The comma-delimited binding record format remains unchanged. A description or key containing a comma already breaks the row for every bind kind, and none of the shipped defaults has one. Reworking the record framing would also require changes to parse_binding_records, static_bindings, and every test that reads rendered rows, so that belongs in a separate change.