fix(web,mobile): a turn you expanded stays open when the next run starts#23608

Problem

When you open a turn's "Worked for …" fold to read its earlier messages, it closes again as soon as the next run starts. The text you were reading disappears. Steps: let a multi-message turn finish, open its "Worked for" fold, send a follow-up. The fold snaps shut about a second later. It also happens if you switch to another thread and back before sending. Mobile's thread feed does the same.

The cause is the effect that watches the latest run (web MessagesTimeline.tsx, mobile ThreadFeed.tsx). It exists to fold away the expansion it adds itself when a run is interrupted. In practice it removed the previous run from the expanded set on every run change, including folds the user opened.

Change

• Web and mobile now remember the one run the interrupt branch auto-expanded.
• When the next run starts, only that run is folded again.
• A fold opened by the user, by a citation or by find stays open.
• A run that was already open when it was interrupted is not marked, so it stays open too.
• If the user toggles the auto-expanded run themselves, the marker is cleared and their choice wins.
• On web, the marker is cleared when the thread changes. Mobile remounts the feed per thread.

Scope and approval

Fixes #17382 GitCafe: https://git.cafe/pingdotgg/t3code/issues/17382 Triaged bug via the automated issue flow.

Verification

Tests

• New browser tests in apps/web/src/components/chat/MessagesTimeline.browser.test.tsx:
  • keeps a turn the user expanded open when the next run starts. It fails on main with expected 'false' to be 'true' on the fold's aria-expanded.
  • keeps a restored expanded turn open when the next run starts. It seeds the remembered fold that a thread switch restores, then starts the next run. It fails on main with the same assertion.
  • folds an interrupted turn it expanded once the next run starts. It guards the interrupt behaviour.
  • keeps an interrupted turn that was already expanded open when the next run starts. It fails on the previous head of this branch.
• cd apps/web && vp test run --project browser src/components/chat/MessagesTimeline.browser.test.tsx: 14 passed.
• vp test run apps/web/src/components/chat/MessagesTimeline.logic.test.ts: 81 passed.
• vp lint on the changed files reports no errors; the warnings are on lines this change does not touch.
• Web and mobile typecheck exit 0.

Evidence

Each pair below is the same scripted interaction in the real web app:

• Before is main at 8d3b061121; after is this branch at c29921141a.
• Dev server against a seeded copy of real data, with a live Claude Opus 5.5 thread.
• Chromium in the T3 preview browser, 1440×900, DPR 1, light theme, no device emulation.
• The change only affects which folds survive a run change, not markup or styles, so viewport and theme variants show nothing extra.

What else changed:

• Folds opened by a citation or by find now also stay open when the next run starts. Before, the latest of them was closed.
• After an interrupt, switching threads and back clears the auto-expansion marker on web. In that case the interrupted run stays open on the next run instead of folding. In the real app an interrupted run carries its interrupt item and is not folded at all, so neither case shows a visible difference there; the browser test covers the interrupt path.

Not checked:

• The mobile change in apps/mobile/src/features/threads/ThreadFeed.tsx was not captured on the iOS simulator.
• ThreadFeed has no render test seam, so mobile is covered by reading it side by side with the web change.
• The desktop app wraps the same web timeline and was not run separately.

Model and harness: Claude Opus 5.5 (with Claude Sonnet 5.5 as a second reviewer) in Claude Code, driven by T3 Code.

🤖 Generated with Claude Code

Review round (head 8ab4001)

Two independent read-only reviews of the full diff against main:

• Claude Sonnet 5.5: VERDICT: MERGE. No blocking findings. Checked the user, citation, find, interrupt, already-open-interrupt and toggle paths on web and mobile. Mobile remounts per thread via RenderErrorBoundary key={selectedThreadKey} in ThreadDetailScreen.tsx. No marker leak across threads, and nothing outside apps/web and apps/mobile.
• Claude Opus 5.5: VERDICT: MERGE. Two low findings and a nit, none blocking:
  • On web, after interrupt → switch thread → switch back, the auto-expanded run is restored open but its marker is cleared. That run stays open on the next run. Mobile resets by remount instead. This is already noted above and has no visible effect, because interrupted runs never fold in the real app.
  • A citation or find that targets the run the interrupt already auto-expanded doesn't clear the marker.

Evidence

Supersedes the c29921141a recordings in the first two rows of the table above. Those predate 8ab4001. The after-only interrupt recording is kept, because interrupted runs never show a fold in the real app.

Setup:

• Before is main 8d3b061. Both changed files are identical to the merge base ed3f965e07.
• After is this head 8ab4001.
• Both are real web dev servers with isolated seeded homes, recorded in the T3 preview browser (Chromium, DPR 2).
• Each row is a fresh thread with a live Claude Haiku 5.5 turn that folds under "Worked for …". Every row clicks that fold open, does the row's trigger, then sends a follow-up.
• Driver: driver-23608.md, sha256 a2a5fe5d9bbe6156c4e0262eb47988263fcee8b9ecdc9bcb0293a231c804a3a0.
• Matrix: proof-matrix-23608.md. Narrated capture report: capture-report.md.

Row 1, rebuilt independently by the second auditor. Left to right: before with the fold open, before after the follow-up (collapsed), after with the fold open, after after the follow-up (still open):

Row 1 rebuild: before open, before collapsed after follow-up, after open, after still open after follow-up

What these recordings don't cover:

• Interrupt rows (interrupt auto-expand, then next run; already-open interrupted run): no visible difference in the real app, because a run with an interrupt item never folds (MessagesTimeline.logic.ts, mobile threadActivity.ts). They are covered by the browser tests folds an interrupted turn it expanded once the next run starts and keeps an interrupted turn that was already expanded open when the next run starts.
• Mobile: unverified. There is no static fixture mode: EXPO_PUBLIC_SHOWCASE needs three real paired servers and cannot start a run. Device access is off on the capture host. Only an EAS dev build plus Metro tunnel with a live provider could show it. The mobile change is covered by both reviews reading it against web.

Proof audit pass 2: VERDICT: SUFFICIENT. Every recording was checked frame by frame, SHAs and the driver hash were checked, and row 1 was rebuilt. Report: audit2-report-23608.md.