fix(desktop): fit browser previews at any window zoom

[incognitojam]

Note

Fixes cropping and wrong click locations in fitted desktop browser previews across independent window and page zoom.

Fitted browser previews crop the page when the desktop window or preview page is zoomed out and leave unused space when zoomed in. The page can also measure a different CSS viewport from the requested size. Screenshots capture the same incorrect surface, and agent clicks miss their targets because their logical coordinates are sent directly to the scaled surface.

Convert requested CSS dimensions to CDP device-independent dimensions using page zoom, and account for host window zoom in presentation scaling. Map agent click coordinates into the fitted surface while preserving logical cursor coordinates. Clicks use the scale successfully applied to their CDP session while a resize or window-zoom update waits. Restore every tab's independent page zoom before refreshing fitted presentations concurrently, so a busy tab cannot delay restoration for the others. Refreshes read the latest requested viewport without rewriting it and check session ownership, preserving queued resizes and preventing a detached session from changing its replacement.

Validation

• Reproduced the cropping in the built-in Browser panel. Exercised the real preview manager in an isolated Electron window using resolveBrowserViewportLayout to derive presentation. All nine combinations of window zoom 80%/100%/120% and page zoom 90%/100%/110% preserve the requested 1280×800 CSS viewport, fill the captured surface to all four edges, and deliver locator clicks to the button. Window-zoom changes in this matrix exercise reapplyZoom without a renderer presentation resend.
• Exercised a 375×667 viewport, a mini-player-sized surface, restoration to panel size, saved screenshot artifacts, and clearing viewport emulation. Full packaged-app verification requires a desktop rebuild.
• Deterministic concurrency regressions cover fill-to-fitted and fitted-to-fill resizes queued during a click, window zoom changing during a click, a busy fitted tab while another tab's page zoom is restored, and a resize queued during recording startup followed by a zoom refresh.
• Queued-action tests verify input coordinates, ordering, and the final native presentation. A second fitted tab refreshes while the first is busy, then the first refreshes when released. A detached-session regression fails when its session-ownership guard is removed.
• All 168 focused tests pass across Manager.test.ts, DesktopWindow.test.ts, and hostedBrowserWebviewStyle.test.ts. Desktop typecheck, targeted lint, and formatting checks pass.

Known limitation: presets rendered at fit scale 1 still inherit window zoom in their measured CSS size. This pre-existing natural-size path is tracked for upstream intake; its guest-zoom composition needs adaptation to the fork's CDP presentation path.

Before / after

The same synthetic 1280×800 page in a fitted surface at 80% window zoom, captured through the real manager on main and this branch:

Before	After

The page-zoom correction at 80% window zoom and 90% page zoom, using renderer-derived presentation. Before, the requested 1280×800 viewport measures 1422×888 and crops the bottom-right edge; after, it measures 1280×800 and fills the surface:

Before page-zoom correction	After page-zoom correction

---

Written by an agent (Codex, gpt-6.1-sol).

[github-actions]

Upstream divergence

Both sides are compared with upstream at the in-order tip 33b7d7b3ae. Excludes .repos/ and pnpm-lock.yaml.

Measure	Base	This PR	Change
Upstream files the fork changes	1,368	1,368	0
Lines changed in them	62,408	62,897	+489
Upstream files the fork removes	41	41	0
Fork-only files	435	435	0
Fork-only lines	74,632	74,636	+4

Upstream files whose difference changes (2)

File	Change	Now differs by
apps/desktop/src/preview/Manager.test.ts	+429	1,065 lines
apps/desktop/src/preview/Manager.ts	+60	578 lines

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	4.9 KiB	5.0 KiB	+139 B (+2.8%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.7 KiB	3.8 KiB	+81 B (+2.1%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.1 KiB	1.2 KiB	+58 B (+5.0%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	20.4 KiB	20.9 KiB	+482 B (+2.3%)	29.3 KiB	✅
Codex	Live turn messages	1	2	+1 (+100.0%)	8	✅
Claude	Total thread wire	4.9 KiB	5.0 KiB	+100 B (+2.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.7 KiB	3.8 KiB	+82 B (+2.2%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.2 KiB	1.2 KiB	+18 B (+1.5%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	20.8 KiB	21.2 KiB	+441 B (+2.1%)	29.3 KiB	✅
Claude	Live turn messages	2	2	0 (0.0%)	8	✅

Baseline: a1861e2 · PR result: 60856bb · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 108.5 KiB
• Claude decoded thread snapshot: 108.8 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.