fix(tui): stop the scrollbar from blanking the transcript on resize

[ericleepi314]

Resizing the terminal blanked the whole transcript — composer and footer kept painting, the transcript region went empty, and it never came back. Reported in the VS Code integrated terminal.

Root cause

TranscriptScrollbar is a row sibling of the transcript ScrollBox and renders exactly viewportHeight rows tall — and that value is last frame's scrollViewportHeight, written by render-node-to-output at paint time.

A flex row stretches its children to the tallest one, so a stale-tall bar stretches the ScrollBox; its inner wrapper (flexGrow:1) fills the extra with blank rows; and the stretched height becomes the next frame's scrollViewportHeight, which re-renders the bar at that same height. A stable fixed point, not a transient — which is why it never recovers.

Widening re-wraps every row shorter, so the gap opens exactly then. Measured on a PTY at 60 → 110 columns: content fell 161 → 102 rows while the ScrollBox stayed at 161. Those ~59 blank rows push the transcript above the terminal viewport, where the resize repaint's ERASE_SCROLLBACK (log-update fullReset → clearTerminal) wipes it from scrollback.

Fix

if (!vp) → if (!vp || INLINE_MODE) in both scrollbars.

Inline mode is the default and has no constrained-height root, so the ScrollBox always fits its content and total <= vp holds — the bar only ever took its !scrollable branch and painted a column of invisible spaces, 40 rows of them. The 1-column gutter stays, so transcriptPanelWidth's reservation still matches what's drawn. Fullscreen pins the root to terminalRows, can't drift, and the bar is real UI there — untouched.

OverlayScrollbar (agentsOverlay.tsx) has the identical shape and the same bug, so it gets the same guard. Its minHeight={0} doesn't help: that permits shrinking below content, it doesn't stop cross-axis stretch.

Ruled out by experiment

Recorded so nobody retries them:

• Recursively marking every Yoga node dirty before calculateLayout — so this is not a layout-cache bug.
• minHeight={0} on the ScrollBox.
• minHeight={0} on the ScrollBox plus both ancestor flex boxes.
• Keying the ScrollBox on cols to force a remount. It fixes the symptom but orphans three useSyncExternalStore subscriptions on the dead handle and cold-starts the virtualizer. The three call sites now carry a comment explaining why their stable deps are only safe while the ScrollBox never remounts.

Verification

• PTY + pyte against the real binary: blank rows after a 60 → 110 resize went 20-22/30 → 7-8/30, holding for a ~35-row transcript. Only widening exhibits the bug; narrowing grows content past the stale-short bar, so there's no gap to open.
• Regression test asserts ROWS, not glyphs — the inline case paints spaces, so "draws no bar characters" passes against the bug. Self-calibrates against a bar-free render, and waits for a settled frame rather than a fixed timeout (which would fail open, since with the fix the bar paints nothing). Mutation-tested: reverting the guard fails 2 of its 3 cases.
• Suite: 5 consecutive full runs at 8 failed / 1608 passed. main's baseline is 8 failed / 1605 passed — same 8 pre-existing failures, +3 from this PR, no flake.
• tsc --noEmit clean, eslint clean on all changed files.

Follow-up, not in this PR

ERASE_SCROLLBACK (ESC[3J) fires on every resize (log-update.ts:188 → clearTerminal.ts:63). In inline mode a full repaint needs only ESC[2J + cursor home; the ESC[3J additionally destroys the user's pre-TUI shell scrollback and every transcript row that scrolled off, on every window drag. It survives this fix — it's what upgraded "scrolled off screen" to "unrecoverable".

🤖 Generated with Claude Code

[github-actions]

Test Results

    1 files      1 suites   7m 59s ⏱️
9 376 tests 9 370 ✅ 6 💤 0 ❌
9 622 runs  9 616 ✅ 6 💤 0 ❌

Results for commit df7f823.