Wide minimum-width pages can be clipped when panels open. Simply asking for a wider window makes code/result comparison difficult on a laptop. Northstar fits actual remaining web space while retaining the document and draft. Behind the scale are learned width, tab lifetime and input-coordinate decisions.
Measure actual web space rather than the whole window
The calculation uses the actual WebContentsView width after code, right panel and split views occupy their space. Scale is the content-width ratio rounded down to one percent, capped at one hundred percent and floored at twenty-five percent. A wide outer window does not imply a wide central web area.
The lower limit means extremely wide tables or canvases can retain horizontal scrolling in narrow splits. This is no mobile-CSS redesign. Responsive pages that already fit keep their layout. Shrinking every page would reduce legibility in the name of readability. The goal and limit are to fit observed overflow, not universally miniaturize content.
Explore the implementation
Reviewed original scale calculation
return Math.max(0.25, Math.min(1, Math.floor((viewWidth / naturalWidth) * 100) / 100));Do not relearn a distorted natural width after shrinking
Zooming out increases the CSS viewport, and responsive wrappers can expand again. Using each current scrollWidth as a new natural width can prevent restoration or cause oscillation. PageWidthFit retains width learned from actual overflow and resets it for a new document.
Current ChatGPT measurement contains a bounded structural check so off-canvas portals and hidden regions do not falsely widen its responsive document. Other sites retain the established scrollWidth measurement. That specialization is not a universal CSS analyzer detecting every hidden element. The same ratio has different meaning depending on which geometry supplied its denominator.
Do not attach a late measurement to a changed page
Before measurement, the code captures URL, documentEpoch and native view width, then reads geometry with a fixed isolated-world function. On return it discards a result if URL, generation, width or page availability changed. A delayed reply during resizing should not apply the old scale to a new layout.
Requests while busy set again and are rescheduled afterward. Short timers coalesce work without repeatedly probing at one hundred percent zoom. The implementation neither rewrites DOM wrapping nor reloads the page. That keeps resizing distinct from editing and limits changes to the document lifetime while a draft is in progress.
A manual scale takes precedence over automatic fit
Explicit zoom switches that tab to manual mode and restores both mode and scale after restart. Even same-origin tabs isolate zoom: one may hold a fixed-width editor while another is a responsive reference. A separate control resumes automatic fitting.
Existing end-to-end tests compare draft value, focus and timeOrigin and check an actual broker click after scaling. That goes beyond fitting a screenshot. Coordinates and input must still identify the same target. Owned fixtures and historical public-site evidence retain separate scopes; their results are not compatibility guarantees for every editor.
Check that a person can continue, not only see the edge
The historical Naver check observed thirty-five percent fitting in a narrow window and restoration to one hundred percent when widened, retaining the document and session. Current specialized measurement and that historical observation are evidence from different points in time. I ran no new live-site test while writing this article.
The displayed code is the core of the reviewed ratio function. The independent preservation module compares synthetic fingerprints only; it reads no manuscript and changes no zoom. The design criterion includes whether the person can continue typing at the same place, not merely see the right edge. Layout convenience must preserve document continuity and manual choice to become a useful workbench capability.