cmux-next agent pane: no blank edge while flinging the transcript

[teamleaderleo]

Summary

While you flung the React agent pane's transcript, every frame showed a blank strip at the edge. In the #16231 bench, all 240 of 240 frames were blank. Two things caused it:

• VirtualTranscript committed a scroll one animation frame after the scroll event.
• It mounted only the viewport plus four rows of overscan, so the offset the next frame reached was never mounted.

Now the scroll handler commits from the scroll event with flushSync. The mounted range also leads in the scroll direction by two scroll steps, capped at four viewports, so the rows a fling reaches next are already in the DOM. When scrolling stops, the lead stays at the last step's size until the next scroll.

Part of #16424. Builds on #16452 (merged). This PR's commits are 16463ee (test) and 3b0dc6c (fix), plus a merge of feat-cmux-next.

Verification

• transcript.test.tsx gets a jsdom test that scrolls up a viewport and a half per scroll event and expects the rows two steps ahead to be mounted right away:
  • Red at the test-only commit: min mounted top 8008, expected <= 5500, because rows stayed at the old offset.
  • Green at head.
• In webviews/: bun run test (280 pass), typecheck and lint:ci pass. build-agent-pane-web.sh --check and build-webviews-app.sh --check pass.
• In-app, tagged build nx-scroll-v1 (fleet job f5bbdb3b, head 3b0dc6c): seeded 5,000 rows, then ran four 3 s flings. The display is 160 Hz and the page runs at 80 Hz, as in the earlier benches.

per fling	before (#16425)	after
Blank edge frames	240–241 of 240–241	1 of 240–241 (the first frame, before any scroll delta)
Frame interval p50 / p95	12–13 / 13 ms	12–13 / 13 ms
Dropped frames	0, 1, 0, 0	0, 1, 0, 0
React render→commit p50 / p95 / max	1 / 1 / 2–4 ms	1 / 1–2 / 1–4 ms
Layout	0 ms	0 ms

The 80 Hz page cap remains (#16424).

Changelog

none

🤖 Generated with Claude Code

https://claude.ai/code/session_01RYQHfug1ZVQDp4eWgwVUtD

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

---

Summary by cubic

Fixes the blank edge strip that showed on every frame when flinging the agent pane's transcript, and pulls in the feat-cmux-next lag and reconnect fixes.

Bug Fixes

• Scroll now commits from the scroll event with flushSync, and the mounted range leads in the scroll direction by two scroll steps (capped at four viewports), so the rows a fling reaches next are already in the DOM.
• Lag resync keeps the live summary, queue and permission through the rebuild and replays missed mux events on top.
• A failed lag fetch ends "resyncing" with "failed" instead of staying stuck.
• A reconnect re-attach fetches the events between the old cursor and the attach page; a lag notice before the new attach waits for it.
• close() rejects pending requests, and failed optimistic prompt rows survive a rebuild.
• A height-only shrink that makes fitting rows overflow opens at the latest row once; registering an unchanged renderer no longer re-renders the pane.

^{Written for commit 3d9a69c. Summary will update on new commits.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository: manaflow-ai/cmux/.coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: 21762a95-d369-4141-9f18-4394721ed058

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Autopilot is currently an internal CodeRabbit preview.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[teamleaderleo]

Subagent review at 3b0dc6c: APPROVE.

• Lead direction: correct both ways, and a delta of 0 behaves as before.
• flushSync: called from the scroll event handler, with no React warnings and no re-entrancy; layout-effect scrollTop writes queue their scroll events for the next frame.
• Trackpad cost: browsers coalesce scroll events, so it's at most one synchronous commit per frame.
• Anchoring: extra runs of the anchoring effect are no-ops while the layout is unchanged.
• Test: discriminates both the rAF hop and a missing lead.

Non-blocking, added to #16424:

• no downward-scroll test;
• the lead stays after a big jump until the next scroll;
• the anchor uses the first mounted row rather than the first visible one, which predates this PR and is slightly widened by the lead.

[cubic-dev-ai]

3 issues found across 5 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="webviews/src/agent-session/acpmux/App.tsx">

<violation number="1" location="webviews/src/agent-session/acpmux/App.tsx:128">
P2: `scroll.delta` retains the last fling indefinitely, so later transcript updates keep mounting up to four extra viewports even after scrolling stops. Clear the directional lead after the scroll commit or scope it to the scroll render so virtualization returns to the normal range.</violation>
</file>

<file name="webviews/src/agent-session/acpmux/direct.ts">

<violation number="1" location="webviews/src/agent-session/acpmux/direct.ts:167">
P2: The reconnect path pages `_acpmux/events` past the attach snapshot, but with `replayLiveState=false` the fetched mux events' effects (queue, status, permission) are discarded: `rebuildKeepingLiveState(undefined)` restores the pre-fetch `summary`/`queue`/`pendingPermission` and replays nothing, and `apply()` then drops the eventStream notification for the same seq because `event.seq <= this.lastSeq`. Any `queued`/`queue_removed`/`status`/`permission_*` event that lands between the attach reply and the end of the fetch (the fetch pages to *now*, not to the attach page) is permanently lost from the live state, leaving a stale queue chip or status until an unrelated later event changes it. Pass `replayLiveState` as true (the lag path's behavior) so events after `resumeAfter` are replayed onto the attach snapshot in seq order; the guard `oldest > resumeAfter + 1` is computed only from transcript seqs, so it does not limit the mux events this fetch returns.</violation>

<violation number="2" location="webviews/src/agent-session/acpmux/direct.ts:277">
P2: This branch still publishes a stale session list when selection changes while `_acpmux/watch` is pending. Discard the response (or merge the current selection) before replacing `this.sessions`, otherwise the selected session can disappear from the picker until another watch update arrives.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

Unanchored findings

2 findings could not be posted inline, so cubic included the feedback here instead.

1. webviews/src/agent-session/acpmux/direct.ts:277

P2: This branch still publishes a stale session list when selection changes while _acpmux/watch is pending. Discard the response (or merge the current selection) before replacing this.sessions, otherwise the selected session can disappear from the picker until another watch update arrives.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. At webviews/src/agent-session/acpmux/direct.ts, line 277:

<comment>This branch still publishes a stale session list when selection changes while `_acpmux/watch` is pending. Discard the response (or merge the current selection) before replacing `this.sessions`, otherwise the selected session can disappear from the picker until another watch update arrives.</comment>

<file context>
@@ -236,28 +244,46 @@ export class AcpmuxDirectClient {
-    this.emit("session changed");
+    const missing = this.selectedSessionId !== undefined && !this.sessions.some((session) => session.sessionId === this.selectedSessionId);
+    if (missing && generation === this.selectionGeneration) this.selectFallbackSession("session changed");
+    else this.emit("session changed");
+  }
+
</file context>

2. webviews/src/agent-session/acpmux/direct.ts:167

P2: The reconnect path pages _acpmux/events past the attach snapshot, but with replayLiveState=false the fetched mux events' effects (queue, status, permission) are discarded: rebuildKeepingLiveState(undefined) restores the pre-fetch summary/queue/pendingPermission and replays nothing, and apply() then drops the eventStream notification for the same seq because event.seq <= this.lastSeq. Any queued/queue_removed/status/permission_* event that lands between the attach reply and the end of the fetch (the fetch pages to now, not to the attach page) is permanently lost from the live state, leaving a stale queue chip or status until an unrelated later event changes it. Pass replayLiveState as true (the lag path's behavior) so events after resumeAfter are replayed onto the attach snapshot in seq order; the guard oldest > resumeAfter + 1 is computed only from transcript seqs, so it does not limit the mux events this fetch returns.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. At webviews/src/agent-session/acpmux/direct.ts, line 167:

<comment>The reconnect path pages `_acpmux/events` past the attach snapshot, but with `replayLiveState=false` the fetched mux events' effects (queue, status, permission) are discarded: `rebuildKeepingLiveState(undefined)` restores the pre-fetch `summary`/`queue`/`pendingPermission` and replays nothing, and `apply()` then drops the eventStream notification for the same seq because `event.seq <= this.lastSeq`. Any `queued`/`queue_removed`/`status`/`permission_*` event that lands between the attach reply and the end of the fetch (the fetch pages to *now*, not to the attach page) is permanently lost from the live state, leaving a stale queue chip or status until an unrelated later event changes it. Pass `replayLiveState` as true (the lag path's behavior) so events after `resumeAfter` are replayed onto the attach snapshot in seq order; the guard `oldest > resumeAfter + 1` is computed only from transcript seqs, so it does not limit the mux events this fetch returns.</comment>

<file context>
@@ -154,10 +153,19 @@ export class AcpmuxDirectClient {
+      if (sessionId) {
+        const page = await this.attach(sessionId, generation);
+        const oldest = page.length > 0 ? Math.min(...page.map((event) => event.seq)) : 0;
+        if (resumeAfter > 0 && oldest > resumeAfter + 1) await this.fetchMissedEvents(sessionId, generation, resumeAfter, false);
+      }
       this.hasConnected = true;
</file context>

if (resumeAfter > 0 && oldest > resumeAfter + 1) await this.fetchMissedEvents(sessionId, generation, resumeAfter);

[cubic-dev-ai]

P2: scroll.delta retains the last fling indefinitely, so later transcript updates keep mounting up to four extra viewports even after scrolling stops. Clear the directional lead after the scroll commit or scope it to the scroll render so virtualization returns to the normal range.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. At webviews/src/agent-session/acpmux/App.tsx, line 128:

<comment>`scroll.delta` retains the last fling indefinitely, so later transcript updates keep mounting up to four extra viewports even after scrolling stops. Clear the directional lead after the scroll commit or scope it to the scroll render so virtualization returns to the normal range.</comment>

<file context>
@@ -118,7 +125,8 @@ export function VirtualTranscript({ rows, onToggleActivity, expanded, registry =
   const layout = measured.layout;
   const reportedLayout = useRef<typeof measured | null>(null);
-  const range = visibleLayoutRange(layout, scrollTop, height);
+  const lead = Math.min(Math.abs(scroll.delta) * SCROLL_LEAD_STEPS, height * MAX_SCROLL_LEAD_VIEWPORTS);
+  const range = visibleLayoutRange(layout, scroll.delta < 0 ? scroll.top - lead : scroll.top, height + lead);
   useLayoutEffect(() => {
</file context>

[teamleaderleo]

Merging. 3d9a69c is the approved head 3b0dc6c with feat-cmux-next (#16452) merged in. The conflicts were only this PR's own additions against base context plus the generated bundle, rebuilt with --check passing; webviews tests (280 pass), typecheck and lint pass. Release compile is green. cmux-next swift test is the lane's existing debt (cx-next-swift-tests).