fix(web): chat stays pinned when the terminal drawer opens during a turn#32826

Problem

While an agent turn runs and the chat is following its end, opening the terminal drawer pushes the newest content below the visible area. It slides back only after the drawer stops moving. When idle, or with reduced motion on, the end stays pinned.

During a turn, MessagesTimeline passes an animated maintainScrollAtEnd so streamed text scrolls smoothly. Legend List (3.7.2) has one animated flag for every trigger, so a viewport resize also gets a smooth pin, queued behind any smooth pin already running.

Change

MessagesTimeline handles LegendList's onLayout, which Legend calls after it has measured the new viewport. When the viewport height changes during a turn (the smooth follow config) and the list is at its end, it calls scrollToEnd({ animated: false }), which pre-empts a running smooth pin and keeps follow on.

• The pin runs in a microtask. Legend applies item remeasurements from the same resize delivery after onLayout returns, so a row that grows while the drawer moves is included. flushSync commits the scroll before the frame paints.
• "At its end" is Legend's isFollowingEnd, or, while that is briefly cleared by the scroll-to-end button's settle step, being within the app's existing 40 px at-end band at the old viewport height.
• A reader who has scrolled away is not pulled back. Wheel, touch, keys and the scrollbar turn the app's follow off, and the handler then does nothing.
• Content growth with a still viewport keeps scrolling smoothly. Width-only changes, idle, reduced motion, Find, citations, position restore, anchored turns and fullscreen apps take the existing paths.

No dependency patch: the fix lives in the app.

Scope and approval

Fixes pingdotgg/t3code#32778

The issue was triaged as a bug through the automated issue flow. Web and desktop share this component; the native mobile app uses its own feed and is unaffected.

Verification

• MessagesTimeline.browser.test.tsx, case "pins the live end at once when the viewport shrinks during a turn": a running smooth pin is replaced by an instant one; a row growing in the same resize delivery is included (191 px short without the microtask); a resize during the scroll-to-end settle still pins (50 px short without the band check); a reader scrolled to the top is not moved. The case fails on main's component. 55/55 tests pass; vp run check:changed passes.
• Reviewed in two rounds by two independent reviewers. Round 2 found no code bugs.

Evidence

Captured on f57ff765d8 (after) against merge-base 20c3dd3689 (before). The current head only merges main on top, with no change to these files. Each side ran its own dev server on a fresh database with the same synthetic Replay turns. Environment: Playwright, Chromium 148, Linux, 1440×900, DPR 1, font scale 1, dark, normal motion, frozen clock. Distances are from the end of the timeline scroll container, read in a ResizeObserver that runs after Legend's, which is the position each frame paints with.

An independent rebuild of R01 from the uploaded scripts gave 80 height changes, all at 0 px.

R01 (before, after, Diff0, DiffT at threshold 16).

R01 before, after, Diff0, DiffT

R02: drawer closed and reopened while markdown streams (before, after, Diff0, DiffT). Before, the newest rows are under the composer; after, the newest Thinking row stays in view. Outside the timeline, 0 pixels differ above threshold 16.

R02 before, after, Diff0, DiffT

Recordings: R02 before · after · R09b before · after · R09b sheet (overlapping return-and-resize variant: 464 px → 0 px).

Every row with its narrated steps, recordings and ResizeObserver series: Evidence-r2.md and, for the R02 and R09b recaptures with matched sidebars, Evidence-r2b.md. Matrix: proof-audit-pass1.md. Seed, drivers and sampler: scripts · reproduction. Earlier rounds (R05, R06, G01 guest page, R08 width-only, R09a released reader): Evidence.md.

What else differs between the sides, outside this change: streaming timing, and anti-aliasing pixels at the composer edge. In R01 the before frames show the newest rows hidden under the composer, which is the bug.

Not verified:

• While the viewport is still, streamed growth follows smoothly and can briefly trail the end (up to 112 px in the R02 and R09b recordings, then decaying), as on main.
• Windows desktop, the reporter's platform. The desktop app renders this same web code; no Electron capture was made.
• Settling, position restore, anchored turns, citations, fullscreen apps and a mounted empty list skip the new handler through the unchanged maintainScrollAtEnd guards; they were not captured.

AI-assisted; verified with the tests and tools listed above.