[mahanti]

Holding message text previews a reply in the existing composer across Messages, Activity conversations, Bestie and conversation tabs. The preview appears after 100 ms; composer focus waits for pointer release. Dragging to select, scrolling, or Escape cancels the provisional target without losing the draft. Reply actions retain the designer-approved MessageIcon, with the icon contract documented separately from thread tabs and links.

Cancellation restores prior following only when no newer reading intent or arrivals supersede the preview. Channel and thread owners preserve the selection anchor when messages arrive during a hold, including partially loaded threads. Explicit Solo replies create threads, and the exact thread reader retains its existing verified-evidence send gates.

The final scroll repairs keep upward channel input from restarting near-bottom following and let a downward wheel at the actual thread end resume following, including wheel input forwarded synchronously by the floating action bar. The thread-scroll measurements and keyboard-preference journey open real thread links. Main through 15fdd36 is integrated. Channel resize restoration also retires an anchor when newer native upward scrolling has superseded its captured position, preventing a queued callback from reversing a continuous wheel gesture.

Validation

• Earlier coverage added three arrivals-during-preview cases (channel, thread, and partially loaded thread), with none removed. Those failed before their repairs in both engines. All six isolated thread-scroll measurement samples passed; corpus, gesture distances, and thresholds are unchanged. The related community-menu test now checks setup focus before opening rather than racing autofocus; restoration and removed-composer assertions remain intact.

• The existing channel scroll case remains unchanged. One additional slow-scroll case splits the same 350px input into 20px and 330px, proving the first upward step survives preview cancellation. Before repair it failed with a zero bottom gap in Chromium and WebKit.

• The existing rejected-Join journey now deliberately returns to the end through the floating message-action bar, after scrolling far enough upward to pause following. It checks actual pointer targets and waits for native gesture completion. Before repair Chromium reproduced the exact hosted overlap of 23.90625px. The notice-overlap and end-position assertions are unchanged.

• Native selection, scroll geometry, top-layer wheel forwarding, and real resize ordering require browser coverage. No assertions were relaxed, no retries were added, and no production test seam or new lifecycle owner was introduced.

• At 321e431, all 114 cases across six affected browser files passed in Chromium and WebKit locally, and all 114 passed on Linux. Linux also passed 63/63 WebKit reply repetitions and 70/70 Chromium search repetitions (seven full runs each, two workers, zero retries), plus 2/2 focused WebKit cases. Final head 50575c6 passed all mandatory push hooks, including types, related unit tests, and design checks. Hosted CI run 42163080324 passed on final head 50575c6: 73 successful checks, one configured Windows skip, no failures. Independent review found no blockers. Human acceptance covered those channel and thread repairs in the Mac desktop app. A subsequent pre-push unit regression found completed tiny near-bottom gestures retained active input intent; 50575c6 clears that intent before the final scrollend sample. All 137 existing owner tests and all 68 affected browser cases pass on Mac and Linux after that one-line repair. All 63 final WebKit reply repetitions passed on Linux with zero retries. Human acceptance of this last boundary remains pending.

The shared composer, ReplyContext, message action bar, and existing channel/thread scroll owners remain responsible for the UI. Earlier desktop checks covered identity/community loading, Bestie/Activity replies, drafts, Escape, and light/dark narrow/wide presentation. This follow-up changes scroll behavior; artwork is unchanged. Ready for review at Arjun’s explicit request on 2026-11-06. Automated validation and independent review are complete. Human acceptance of the final completed-tiny-gesture boundary remains pending; this readiness request does not attest that check was performed.

Conflict and shallow-scroll follow-up (2026-11-06)

Integrated main through 15fdd36 without rewriting the PR history. The MessageRow conflicts retain reply holds and incoming agent activity behavior/tests. The community-menu test retains both setup-focus checking and awaited menu autofocus. The later authorization-fixture overlap retains main’s keyboard option-focus sequence and all custody/recovery assertions. Removed the duplicate ChannelTimelineProps import created by the automatic merge.

Kalvin's P3 coverage question exposed a concrete Chromium boundary: a restored row settled at Y=54.984375 for a requested Y=55.984375. Exact equality kept the old anchor active and prevented arrivals from resuming near-bottom following. The existing channel owner now accepts its one-pixel scroll tolerance when retiring that settled standalone anchor; same-row, membership, and bottom-clamp guards remain intact. No new lifecycle state, timer, retry, or scroll policy was added.

Added two browser scenarios, with none removed: channel and thread arrivals after a completed 20px upward gesture, before further scrolling. Native wheel completion and real layout/delivery require the browser layer. The channel case failed before repair in Chromium. Channels retain their existing <80px follow policy; threads retain the upward reading intent already implemented in this PR. The tests document and exercise that distinction.

The current Inbox fixture uses main's event-based reading-position readiness and cleanup. This supersedes the earlier PR fixture synchronization while retaining every unread assertion and the existing fake-time limits. The resolved complete file passed 150/150 tests locally.

Final synchronization retains main's thread catch-up test fix from #2837, including its measurement frames and the negative assertion at 299ms before catch-up at 300ms. The earlier local synchronization change is superseded by that upstream version; the PR's Solo thread composer assertions remain intact.

The Linux WebKit failure on 17ee960 exposed two resize/scrollend orderings in the existing channel owner: a queued no-op anchor restoration could miss the final follow sample, and a late native wheel step could be reclassified as active input after the gesture had ended. The repair samples a fresh converged restoration after reader input while retaining preview-owned and cold-restoration anchors, and preserves completed-gesture state when superseding stale restoration. Production changes are eight additions and one replacement in ChannelTimeline.tsx; no tests, thresholds, timeouts, or retries changed for this repair.

Validation of that exact source snapshot on Linux: 292/292 owner tests; 34/34 complete message-click-reply and thread-resize browser cases across Chromium and WebKit; 30/30 focused shallow/slow/pending-arrival repetitions with zero retries. After integrating main, aa55728 passed all 58 cases across message-click-reply, thread-resize, activity-lifecycle and local-agent-activity in both engines on macOS (4.1 minutes). Independent review found no blockers (9/10). All mandatory push hooks passed on aa55728: TypeScript and related unit tests (498.56 seconds), design types/guards, Rust selection, and policy checks. DCO passed, and GitHub reports the PR mergeable. Hosted CI run 42271653744, attempt 2 passed on aa55728. Both required checks (CI required and DCO Check) passed, and GitHub reports the current head mergeable. All browser journey and repeated-check shards passed in the first attempt. The native lane exceeded its existing 20-minute budget, also reproduced on current main; a targeted retry of that cancelled lane and the aggregate passed without changing the budget or coverage. Native human acceptance remains separate.

The owner returned the PR to ready-for-review at 17:15:48 UTC on 2026-11-06; that concurrent status update is preserved. Native human acceptance of this precise boundary remains pending, and no review-completed attestation is added. In the updated desktop build, hold message prose until the reply preview appears, scroll upward about 20px and release, then observe a peer arrival: a channel near its end follows it, while a thread preserves the reading position. The draft and canceled reply state must remain intact. Continue farther upward to verify ordinary paused reading. Hosted checks and this attended desktop check are tracked separately from the macOS fixture results above.

[mahanti]

🤖 Independent agent review of 92ba37af9 found no blockers in the reply repairs and incoming forwarding/Inbox integration (9/10, bounded source review). The five affected browser files passed all 96 Chromium/WebKit cases, including the four formerly failing WebKit cases. Arjun explicitly accepted identity/community loading and the requested native reply flows on 2026-11-05. Chromium visual inspection also covered dark-wide and light/dark narrow layouts at 200% interface size. These are separate from the agent’s browser behavior checks.

These captures show actual updated components with disposable synthetic fixture data. They establish the displayed light-theme browser states only.

Reply controls at narrow width: the Reply message icon appears beside the preserved Forward action.

Inline reply: a 100 ms hold presents the preview inside the existing composer and retains the draft.

Escape: canceling removes the reply target while preserving the draft and reading position.

Required hosted CI and DCO passed on 92ba37af9: https://github.com/block/buzz-app/actions/runs/42082454184. Human acceptance and independent agent review are complete. The PR is ready for review, with re-review requested from kalvinnchau; code-owner approval is still required before merge.

[kalvinnchau]

🤖 Two behavioral regressions and one memoization issue found. The held-pointer CI failure remains unresolved despite passing local repetitions.

[kalvinnchau]

🤖 Two additional findings: delayed text selection loses focus, and Reply icon guidance contradicts the implementation.

[kalvinnchau]

🤖 One remaining cancellation defect: withdrawing the reply preview does not restore live-bottom following.

[mahanti]

🤖 Addressed the P1 thread-scroll measurement failure in a6741b37c7dc19bea044627a44f38538de3798cf, published in final head d21eda9596255d2e9fcb815bc1c8e12f4bd4be63.

Both opening paths now use a real Buzz thread link instead of inline Reply. The opening click remains inside the measured interval; scrolling waits for the opening pane's geometry to settle before aiming input. Corpus, gesture distances, and assertion thresholds are unchanged. All six isolated open/fling/reverse samples passed across Chromium and WebKit.

The P2 selection-anchor fix and its channel/thread/partially-loaded-thread regressions are included. All 114 affected browser cases passed, including the notification and cancellation cases that failed in prior hosted WebKit CI. Mandatory hooks and DCO passed; independent review and Arjun's integrated desktop acceptance are complete. The new hosted run is still pending: https://github.com/block/buzz-app/actions/runs/42132643607. I am keeping the PR draft until that result is checked.

[mahanti]

🤖 Fixed and pushed as 50575c6. Hosted CI run 42163080324 completed successfully: 73 checks passed, one configured Windows skip, no failures. Both previously failing jobs are green. All review threads are resolved. Mandatory push hooks passed; final repair passed 137 owner tests and 68 browser cases on Mac and Linux, plus 63 Linux WebKit reply repetitions without retries. The PR remains draft pending human acceptance of the final tiny-scroll/scrollend boundary; earlier Mac checks were accepted. https://github.com/block/buzz-app/actions/runs/42163080324

[kalvinnchau]

🤖 One remaining WebKit reading-position failure; the prior native-selection and thread-scroll measurement issues are repaired. Required CI remains red.

[kalvinnchau]

🤖 No new proven regression found. The completed tiny-scroll contract still needs explicit acceptance and coverage; see inline comment.

[kalvinnchau]

🤖 [P3] Cover arrivals after the completed shallow scroll

The slow-scroll case checks a positive gap after 20 px, then scrolls another 330 px before testing arrivals. It therefore does not distinguish the completed shallow-gesture contract: ChannelTimeline.tsx:1824 clears upwardGesture before final position sampling and can resume following within its 80 px bottom threshold, while ThreadPanel.tsx:1879 keeps a shallow upward move paused. The thread variant is skipped here. This is a contract/coverage question, not a proven regression against main. Please explicitly accept or reconcile the two behaviors and assert an arrival after the shallow gesture has completed, before the second scroll; cover the thread behavior or document why it differs.