fix(web): markdown previews keep their scroll position across tabs#23503

Problem

Switching from a scrolled rendered Markdown file to another file tab and back resets the preview to the top. A long document at 1,800 px returns at 0 px.

Change

Remember a bounded, session-local scroll offset per environment, workspace, and file. Restore it when the rendered preview remounts, retry while initially short content grows, and yield to user input. Save again during teardown so an immediate tab switch retains the latest offset.

Scope and approval

Fixes #22300 GitCafe: https://git.cafe/pingdotgg/t3code/issues/22300 Triaged bug via the automated issue flow. Changes affect rendered Markdown in the shared web/desktop file preview panel.

Verification

Bot follow-up: ab2b11d386 applies the test layout class and resets its hoisted mock through setup. The browser regression passes (1/1), scoped formatting/lint pass, fresh web typecheck passes, and the web production build passes. This changes only the browser test; production and fixture bytes remain identical to the captured revision.

CI follow-up: d2f01ac39f registers the directly invoked fixture seed as a Knip entry point. Knip files/dependencies and exports checks pass (2/2), and scoped formatting/lint pass. This follow-up changes only knip.jsonc; production, seed script, and fixture bytes remain identical to captured c133071f33. Evidence columns retain that actual capture SHA.

• The browser regression fails on the unfixed panel with expected 0 to be 1800 and passes with the fix.
• Focused file-preview tests: 14 passed. Targeted lint and web typecheck passed after merging main.
• Two independent final code reviews returned MERGE; the proof audit returned SUFFICIENT. The independent F1 rebuild restored 1,800 px with a pixel-identical Markdown body; 430 toolbar-background pixels differed by at most 8 RGB levels (0 above threshold).
• The lead repeated the file-tab sequence in the real development app: before returned 0 px; after restored 1,800 px.

Full evidence index, every recording and narrated timestamp. Fixture matrix.

Before recording · After recording.

Capture environment

The preview screenshot exporter caps wide PNG output at 1280 pixels; requested CSS 1440×900 is achieved, while PNG export is downsampled 8/9 relative to CSS and 4/9 relative to native device pixels. Native-resolution wide still PNG is cannot-verify with this tool. Recordings retain 2880×1800. Every paired PNG dimension was checked with PIL, and all 40 videos were probed with ffprobe; dimensions are not inferred from preview_resize requests. The wide export cap is present in apps/server/src/preview/ServerBrowserPage.ts (MAX_SCREENSHOT_WIDTH).

Browser: Chrome154.0.0.0, Linux X11 x86_64, T3 collaborative preview; two dedicated agent tabs with explicit tabId on every call. Date constructor/Date.now frozen to 2026-10-22T06:00:00Z; requestAnimationFrame/performance.now/timers remain functional. Each fixture/variant begins with a fresh browser document to clear module cache; within each recording the document and cache persist. Theme comes from preview_set_appearance, with dark class confirmed in recorded environment. document.fonts.status=loaded. Both use the same system font stack: -apple-system, BlinkMacSystemFont, Segoe UI, system-ui, sans-serif. Linux fontconfig resolves sans-serif/monospace to LiberationMono-Regular.ttf, SHA256 395fa5ab8d40c8eba390ced528744ea75a7f69aabf3e68b6f925ca0e39a27370; this is OS resolver evidence, not a CDP per-glyph font assertion. No web font resources or document FontFace entries were loaded.

Lead-owned isolated origins: BEFORE localhost:5807, AFTER localhost:6369. Existing synthetic triage threads and existing apps/server secondary workspaces; another sandbox thread was created through the app command API without starting an agent. No live database writes, no server launched or stopped by this tester, no mocked panel or isolated component harness. Update toast dismissed through its real Dismiss notification button before authoritative sequences.

Panel metrics: wide normal Markdown viewport 539×808 CSS px; narrow sheet 342×752. Explorer narrows these to 283/86px respectively; browser scroll anchoring moves 1200 to 1976.5/5996.5 while preserving the visible content anchor, and hiding explorer returns 1200. Wide maximize expands viewport to 1184px; narrow sheet maximize leaves 342px unchanged. Every checkpoint includes actual dimensions and scroll heights.

Evidence

Whole-app pairs include environmental differences outside the scroll fix: worktree labels, sidebar ages, focus, and explorer state. These are listed below.

All BEFORE columns: 96bae5465b. All AFTER columns: c133071f33. Diff0 is absRGB×8 clipped 255; pixel count uses maxRGB>0. DiffT marks maxRGB>16/255 in magenta. Whole-app counts and fractions include all chrome; the threshold is diagnostic, not an automatic pass.

Representative F1, wide/light, return to alpha:

BEFORE 96bae5465b — resets to 0

BEFORE Markdown return at 0

AFTER c133071f33 — restores 1800

AFTER Markdown return at 1800

What else changed

• Whole-surface worktree names and path breadcrumbs differ (issue-22300 versus issue-22300-current-main), with different truncation. This is retained in the diffs.
• Sidebar thread identities, creation ages, ordering and selected/hover backgrounds differ between isolated databases. Freezing the browser clock does not rewrite their persisted creation times or labels already computed before the hook. These are known chrome noise, not scroll-fix effects.
• The Files explorer shows each revision’s actual repository tree and worktree-specific selection/expansion; the new committed seed files exist only on AFTER. Explorer list differences are preserved. Common Markdown fixture bytes match.
• Source view may have highlight/selection/focus and gutter differences; final checkpoints wait for actual line rendering. Requested line 80 is visibly centered and confirmed by measured bounds. Loading states remain visible in recordings where they occurred.
• Real input leaves focus rings/tab-close hover or selected source/reveal line styles. Narrow sheet shows its real clipped breadcrumb and a sliver of underlying chat/composer; no cropped proof or synthetic mobile UI.
• Intended Markdown body and scrollbar differences occur at restored offsets (BEFORE 0 versus AFTER 1800/900/1200), including clamped retry and later image layout. Pointer/keyboard takeover endpoints agree at 0. Explorer width changes wrap text and invoke normal browser scroll anchoring on both revisions. Wide maximize changes width; narrow maximize is a no-op.
• Scrollbar visibility can differ at a checkpoint because real input and scroll timing differ. DiffT counts include those pixels. No toast obscures the final proof.

These differences prevent a claim of pixel-identical whole-app chrome. Compare control checkpoints (same scroll position), per-row sheets, metrics and recordings; no RGB threshold is treated as an automatic acceptance test.

Unverified surfaces and cases

Native Electron/macOS could not be verified on this Linux VM. Native mobile does not use this component.

Implemented with Claude Opus 5.5 and Codex GPT-6.1-Sol through T3 Code's Claude and Codex harnesses.

Fixes pingdotgg/t3code#22300