Recorder window grows horizontally without bound on fractional display scaling (Windows, 125%)

[yoshihiro-matsumoto]

Summary

On Windows with fractional display scaling (verified at 125% / DPI 120), the recorder HUD widens continuously from the moment it launches, with no user interaction, until it spans the display.

Environment

• Windows 11 x64, display scaling 125% (GetDpiForWindow returns 120)
• Commit 87fdf397ea85f07595b2645bc52cdbcda14eb36f, installed via install.ps1
• Electron 43.1.1, Node 24.19.0

Steps to reproduce

1. Set Windows display scaling to 125%.
2. Launch Skill Recorder.
3. Watch the HUD window. No clicks required.

Observed

Polling GetWindowRect on the HUD window:

elapsed	outer size
0.0s	401 x 500
1.2s	418 x 494
5.0s	507 x 494
20.0s	1306 x 494
45.0s	1403 x 494 (still growing)

Height stays at 494. Only the width grows, at roughly 15-20 px/s.

Analysis

fitRecorderHeight in electron/recorder-window-sizing.ts reads the outer width via getSize() and feeds it straight back into setSize():

const [outerWidth, outerHeight] = win.getSize();
// ...
win.setSize(outerWidth, targetOuterHeight);

Both APIs operate in DIPs. Under fractional scaling the DIP -> physical -> DIP round trip is lossy, so every call widens the window slightly. The width change reflows the renderer, the ResizeObserver in src/Recorder.tsx fires again, and fitRecorderHeight runs again. The result is a self-sustaining rAF-paced loop.

At 100% scaling the round trip is lossless and the height early-return terminates the loop, which is likely why 1a582cc ("fix(ui): prevent recorder window width growth") appeared to resolve this.

Related: the same rounding perturbs the resulting content height, so Math.abs(currentContentHeight - targetContentHeight) < 1 may never hold and the loop never settles on its own.

Suggested fix

Pin the outer width to its first observed value rather than re-reading it on each call, and give the height comparison a small tolerance:

const pinnedOuterWidths = new WeakMap<RecorderWindowSizingTarget, number>();
const HEIGHT_TOLERANCE = 2;

// ...
if (Math.abs(currentContentHeight - targetContentHeight) <= HEIGHT_TOLERANCE) return;

const [outerWidth, outerHeight] = win.getSize();
let pinnedWidth = pinnedOuterWidths.get(win);
if (pinnedWidth === undefined) {
  pinnedWidth = outerWidth;
  pinnedOuterWidths.set(win, pinnedWidth);
}
// ...
win.setSize(pinnedWidth, targetOuterHeight);

With this applied locally the window settles at 402 x 494 and stays there (verified over 45s), and recorder-window-sizing.test.ts still passes.

Note that the existing test double models a lossless DIP mapping, so it cannot catch this class of bug. A case whose setSize simulates fractional-scale rounding would help.

[GiorgioUghini]

This is already fixed in 0.5.0, but yes, thank you for raising this issue!

[GiorgioUghini]

I'll keep this open until 0.5.0 is released tomorrow