Chat can return to the first message instead of the tail when switching back to a populated session

[shanselman]

Summary

A new fixture-backed real-application smoke test intermittently catches a session-return scrolling failure in the production-shaped Release app. After selecting a 240-message session A, selecting session B, and returning to A, the selected conversation is correct but the viewport remains at messages 1-3 instead of the final message.

The screenshot confirms the failed visibility assertion. The exact cause is not established: investigate application scroll/lifecycle behavior and possible UI Automation timing effects before changing production code. Do not weaken the assertion or add a blind retry to obtain a green run.

Related: #1424 (Track WinUI ItemsView fix and retire Reactor preview.12 Markdown workaround) and #1426 (fix(chat): avoid session-switch crashes without clipping Markdown lists).

Reproduction

The new harness currently lives in the local feature/gateway-fixture-harness working tree and has not been published yet. Its files and proof are being handed to the fix agent separately. It launches the real application against an authenticated loopback WebSocket Gateway with synthetic data, not a mock chat provider. No real Gateway, WSL installation, provider, or AI call is required.

1. Build an unpackaged Release app with DevBuild=false.
2. Run scripts\test-gateway-fixture.ps1 -AppPath '<current Release OpenClaw.Tray.WinUI.exe>' -Configuration Release -Screenshots.
3. In GatewayFixtureUiTests.SessionPickerSwitchesRealHistoriesAndShowsMessage240AtBothWidths, select Fixture: 240-message conversation through the actual composer session-picker flyout.
4. Verify its final line, FIXTURE LONG END 240, is naturally visible before issuing any manual scroll command.
5. Select Fixture: research conversation, verify that conversation's content, then select the long session again.
6. On the failing repeat, the long conversation stays at its beginning for the 20-second visibility deadline.

Expected behavior

Returning to the long conversation after it was at the tail should show the actual final message, without requiring a manual scroll repair. Session identity, transcript content, and the user's intended scroll position must remain coherent. Preserve row-targeted navigation and virtualization; checking only the current scroll extent is insufficient.

Observed evidence

• Screenshot shows the connected app, the long session selected, and FIXTURE LONG BEGIN: message 001. followed by messages 002 and 003. The final marker is not visible.
• Gateway trace reports successful handshake, session/model/supporting reads, and successful chat.history for agent:main:main, agent:main:fixture-long, and agent:research:fixture-other. There are no unsupported requests in this run. The return to the long conversation uses already loaded history.
• Latest Release smoke: 22/22 profile/app tests passed; 3/4 native UI tests passed; this natural-tail test failed. Earlier Release runs passed, so the failure is intermittent and earlier successes must not replace the failed result.
• Latest Debug smoke passed 26/26. An earlier Debug run also became unresponsive while inspecting the long transcript; its cause is unconfirmed and may be separate. The harness now bounds UI tests and cleans up owned processes even when a synchronous UI Automation call stalls.
• Required repository build, Shared tests (4,034 passed, 32 skipped), Tray tests (3,006 passed), and WinNode CLI tests (127 passed) passed.

Environment: Windows x64, .NET runtime 10.0.12, Windows App SDK 2.4.0, Reactor preview.12. The failing Release runtime configuration has System.Diagnostics.Debugger.IsSupported=false and no attached debugger.

Proof bundle identifiers for the coordinating maintainer/agent:

• Failing run: 3270bac7f7854ca8a2737baec345be85, including failure.png, run.json, gateway-requests.jsonl, and the redacted app log.
• Final Release TRX bundle: e5404722239f4afa961258b7b5582204.
• Earlier Debug hang: ba6aa16080e340ab98bb98811d9821f6.

Artifacts are retained in the coordinating session and will be supplied to the fix session. These identifiers are not public download links.

Investigation boundaries

Start with the real session-selection path and ReactorItemsViewScrollController / InitialTailPositioner: request-key transitions, layout readiness, cached-session returns, and pending bring-into-view lifetime. Rule out test-driver side effects before assigning the cause.

Keep the fixture harness and the product fix in separate PR scopes. Add focused regression coverage, preserve normal scrolling and scroll-away behavior, and use the populated fixture for native runtime proof when safely available. Do not introduce extent-only scrolling, remove virtualization, replace the real provider, or weaken the 240-message test.

Isolation requirement

Do not use installed pairing data, existing Gateway credentials, or the default WSL distro. Fixture runs require owned profiles, ports and identities plus the explicit host-side-effect isolation guards. Do not launch a main-branch/older app with only OPENCLAW_TRAY_DATA_DIR: loopback WSL keepalive and OS registration paths are not fully isolated by that variable alone.

[clawsweeper]

Codex review: keeping this open for maintainer follow-up; there is still a little grit to resolve. Reviewed September 17, 2026, 5:18 PM ET / 21:18 UTC.

Summary
Keep open: the controlled comparison rejected the proposed fix, and current main contains no subsequent chat-path repair. This collaborator-owned issue remains actionable and has an open investigation PR.

Reproducibility: unclear. for exact current main: a concrete guarded Release A/B/A scenario has reported failures, but this review did not execute it or inspect the retained failure bundle.

Root-cause cluster
Relationship: canonical
Canonical: #1437
Summary: This issue owns the unresolved viewport failure; the open fix PR remains an investigation vehicle, not a verified solution.

Members:

• fixed_by_candidate: fix(chat): restore the tail when switching conversations #1438 - Explicitly linked candidate fix remains open, with its present implementation rejected by controlled testing.
• adjacent_distinct: feat: explore and test the app without a live Gateway #1439 - Provides the isolated real-application fixture and preserves this failure without changing production scrolling.
• adjacent_distinct: Track WinUI ItemsView fix and retire Reactor preview.12 Markdown workaround #1424 - Tracks dependency-workaround retirement; a shared cause with this viewport failure is not established.

Proposal only: this assessment does not dispatch repair, suppress jobs, mutate sibling items, close, or merge anything.

Ways to help us reproduce this

• Add expected vs actual behavior.

Next step
Continue investigation through #1438 (fix(chat): restore the tail when switching conversations); the existing paired PR and no-new-fix-pr label preclude a competing repair.

Review details

Best possible solution:

A lifecycle-correct row-targeted restoration should pass the unchanged guarded A/B/A final-message check while preserving virtualization, streaming follow, and deliberate scroll-away behavior.

Do we have a high-confidence way to reproduce the issue?

Unclear for exact current main: a concrete guarded Release A/B/A scenario has reported failures, but this review did not execute it or inspect the retained failure bundle.

Is this the best way to solve the issue?

No: the session/generation-key candidate was explicitly rejected by the controlled comparison; lifecycle tracing must establish the repair before another production change is accepted.

AGENTS.md: found and applied where relevant.

Remaining risk / open question:

• Application lifecycle behavior and UI Automation timing have not yet been separated as causes.
• The detailed failure bundles remain in the coordinating session; this review inspected the reported comparison, not its raw runtime artifacts.

Codex review notes: model internal, reasoning medium; reviewed against 33839971e315.

Label changes

Label justifications:

• P2: The intermittent viewport failure disrupts returning to a conversation without establishing message loss or a blocked first-use workflow.
• impact:ux-friction: The correct conversation loads, but users can be left at its beginning instead of the expected tail.

Evidence reviewed

What I checked:

• Rejected candidate remains open: The supplied body of fix(chat): restore the tail when switching conversations #1438 (fix(chat): restore the tail when switching conversations) explicitly rejects its current candidate: the unchanged Release scenario passed 4/5 baseline runs and 0/5 candidate runs. A live metadata check confirms the PR remains open, draft, and unmerged. (4101ee5d21e0)
• Current session and tail ownership: The timeline constructs its initial-tail request from session identity, generation, history revision, and scroll token. The controller handles request changes, waits for layout, and calls StartBringItemIntoView for the captured final row. This establishes intended behavior, but does not establish the intermittent failure's cause. (src/OpenClaw.Tray.WinUI/Chat/ReactorItemsViewScrollController.cs:110, 33839971e315)
• Existing scrolling contract: Tests explicitly preserve row-targeted navigation, bottom anchoring, queued-request identity checks, and scroll-away detection while rejecting extent-only ScrollTo and forced UpdateLayout approaches. (tests/OpenClaw.Tray.Tests/ChatTimelinePresentationTests.cs:56, 33839971e315)
• Release-to-main comparison: GitHub's comparison from the supplied v2026.9.4 commit to reviewed main lists three commits affecting automation, packaging, and the Local AI icon, with no chat implementation changes. The latest-release endpoint still identifies v2026.9.4. No fixed release or main-only fix was established. (33839971e315)
• Feature-history routing: Local path history identifies Karen's bottom-anchoring work and earlier Reactor migration. GitHub maps the anchoring commit to karkarl. Local blame and follow-history inspection encountered unavailable historical blobs, so no verified line-introduction claim is made. (src/OpenClaw.Tray.WinUI/Chat/ReactorItemsViewScrollController.cs, ca8fb4a404ad)
• Proof scope: The prepared image was inspected and shows the Markdown wrapping proof from fix(chat): avoid session-switch crashes without clipping Markdown lists #1426 (fix(chat): avoid session-switch crashes without clipping Markdown lists). It does not depict this failure. The issue and investigation describe retained local failure screenshots and traces, but those bundles are not public download links or present in the supplied manifest.

Likely related people:

• karkarl: Suggested for follow-up; no historical authorship or introduction is verified. (role: unverified routing candidate; confidence: low)
• shanselman: Suggested for follow-up; no historical authorship or introduction is verified. (role: unverified routing candidate; confidence: low)

How this review workflow works

• ClawSweeper keeps one durable marker-backed review comment per issue or PR.
• Re-runs edit this comment so the latest verdict, findings, and automation markers stay together instead of adding duplicate bot comments.
• A fresh review can be triggered by eligible @clawsweeper re-review comments, exact-item GitHub events, scheduled/background review runs, or manual workflow dispatch.
• PR/issue authors and users with repository write access can comment @clawsweeper re-review or @clawsweeper re-run on an open PR or issue to request a fresh review only.
• Maintainers can also comment @clawsweeper review to request a fresh review only.
• Fresh-review commands do not start repair, autofix, rebase, CI repair, or automerge.
• Maintainer-only repair and merge flows require explicit commands such as @clawsweeper autofix, @clawsweeper automerge, @clawsweeper fix ci, or @clawsweeper address review.
• Maintainers can comment @clawsweeper explain to ask for more context, or @clawsweeper stop to stop active automation.