resize-pane resizes the pane grid but not the layout rectangle that renders itSeverity: HIGH
Status: ✅ FIXED
Files: src/cmd/cmd.zig:1221, src/layout.zig:173, src/window.zig:387-417, src/server/render.zig:289-351
cmdResizePane computes a target size and then does exactly one thing with it:
// src/cmd/cmd.zig:1221
pane.resizeTerminal(target_w, target_h) catch return .err;
window.invalidateBorderFormat();
return .ok;
The layout tree is untouched. Layout.proportion is assigned in exactly one
place — split creation:
// src/layout.zig:173
.proportion = prop,
Every other reference in the tree is a read (window.zig:405,409,
render.zig:466,486, server.zig:2194,2200,2298,2309, layout.zig:268,274).
There is no setProportion / resize API on Layout.
That matters because the tree, not the pane grid, is the source of truth for geometry:
collectPaneBounds → splitSizes(lw -| 1, s.proportion)) and copies only
pb.w × pb.h cells into the merged screen (render.zig:289-351).Window.resizeNode walks the same tree and calls
pane.resizeTerminal(frame.lw, frame.lh) from splitSizes(...)
(window.zig:387-417).So after resize-pane -R 10:
| Action | Observed effect |
|---|---|
| Grow the pane | Nothing — the renderer still clips to the old tree rectangle |
| Shrink the pane | The tail of the pane goes blank (x < cls.items.len fails → Cell.empty()) |
Any later Window.resize |
Reverts the grid entirely, because it is recomputed from proportion |
server.zig:1179 has the same shape for the resize_left/right/up/down
actions, so those are equally ineffective.
The layout tree is the single source of truth for pane geometry everywhere else
in the codebase, and resize-pane is a user-facing command documented in
README.md and docs/build-run.md. Bugs #220 (swap_pane_* not resizing
panes) and #284 (cmdRotateWindow rotating the pane list but not the tree)
are the same class and were both fixed by making the tree follow — this is the
remaining member of that family.
Layout gained a real resize operation, so the tree — not the pane grid — is
what moves:
// src/layout.zig
pub fn resizePane(self: *Layout, pane: *Pane, dw: i32, dh: i32) bool
resizePane walks towards the target and calls adjustAxis for each requested
axis. adjustAxis prefers the deepest split on the path whose direction
matches, then moves that split’s first child by delta cells (negated when the
target lives on the second side), clamped so both children keep at least one
cell. It writes the proportion back with a half-cell bias —
(want + 0.5) / avail — so splitSizes’ truncation reproduces exactly the
requested size instead of landing a cell short.
Callers translate the request into a delta and let the tree absorb it:
cmdResizePane (src/cmd/cmd.zig) computes dw/dh from the relative
adjustment, or from exact - current for -x/-y, and on success calls
window.resize(...) so every pane is recomputed from the tree.resize_left/right/up/down actions (src/server/server.zig) now route
through one resizePaneBy helper instead of four copy-pasted blocks.The previous direct pane.resizeTerminal path is kept as a fallback for
windows with no split on the requested axis. That is what a single-pane window
is, and it preserves the behaviour the #289 and #391 tests lock in (explicit
-x/-y still sets the grid exactly, huge relative adjustments still saturate
or error). Layout.resizePane returns false in that case, which is how the
caller knows to fall back.
Explicit -x/-y on a multi-pane window is now a delta against the current
size rather than an absolute set; when the pane is a direct child of the
adjusted split the result is exact, and for a deeper pane it is the closest the
single-proportion model can express. That limitation is inherent to representing
a layout as split proportions and matches how tmux behaves.
resize-pane moves the layout tree, not just the pane grid — bug #511
(src/cmd/cmd.zig) splits a window, records both panes’ findPaneBounds
rectangles, runs resize-pane -R 4, and asserts:
Window.resize, which is precisely what the old
grid-only implementation failed: it was recomputed from the unchanged
proportion and reverted.Verified with the full suite on Zig 0.16.0 and 0.17.0: 1043/1043 passing on both.