Uh oh!
There was an error while loading. Please reload this page.
fix(desktop): prevent collapsed workbar flash - #3794
Conversation
hqhq1025
commented
Aug 25, 2026
Visual evidenceSame viewport, first-send state, and fixture backend. BeforeThe collapsed Workbar Suspense fallback temporarily occupies the right side.
AfterThe collapsed Workbar remains absent while the lazy surface resolves.
|
hqhq1025
commented
Aug 25, 2026
CI note: the current failure occurs before the desktop build and is inherited from |
Astro-Han
left a comment
There was a problem hiding this comment.
I reviewed this head and found a blocking issue.
[P1] Claimed queue-test migration not present — build fails
execution-host-queue.test.ts:248/249/271 still references queryTurn/stopTurn/startTurn missing on RuntimeHostConnection. Hosted test is red.
简体中文
存在测试迁移缺失阻断。9114804 to
5f38f4eComparehqhq1025
commented
Aug 25, 2026
Rebased onto current Local verification on
Please re-review the refreshed head. |
Astro-Han
left a comment
There was a problem hiding this comment.
Update on 5f38f4e:
The prior P1 (queue-test still references removed RuntimeHostConnection methods) is now closed. The new head has rebased onto main with #3796's fix, execution-host-queue.test.ts no longer references queryTurn/startTurn/stopTurn (grep 0), and hosted test is now pass (run 32852297860).
The workbar-host fallback logic remains correct (no P0-P3) and is now merge-ready pending human decision.
简体中文
该头 P1 已随 rebase 闭合。5f38f4e to
4fc642aCompareMake the lazy Workbar fallback mirror the resolved surface visibility so creating a session cannot briefly open a collapsed panel. Add Electron coverage that records transient right-panel visibility during the first send. Generated-by: Codex
4fc642a to
5b9627dCompareAstro-Han
left a comment
There was a problem hiding this comment.
The flash is real and the fix is at the right place: the lazy fallback was rendering a right-placement card unconditionally, so a collapsed workbar still flashed a loading panel while WorkbarSurface resolved on the first send.
I checked that the fix does not trade one flash for another: .maka-workbar-workspace-contents is display: contents, the cards position by grid-area, and [data-collapsed] is display: none — so omitting a card and rendering it collapsed are equivalent for layout, and nothing shifts when the surface resolves. The new e2e was verified red on main, which is the part that matters most for a one-frame regression.
Approving. One P3 and one note inline, neither blocking.
AI use: Claude Code assisted with source investigation; the analysis and conclusions are my own.
简体中文
闪烁是真实的，修复位置也对：懒加载的 fallback 无条件渲染了一张 right placement 的卡片，所以首次发送时即使工作栏是收起的，也会闪出 loading 面板。
我确认了这个修复没有用一种闪烁换另一种：.maka-workbar-workspace-contents 是 display: contents，卡片靠 grid-area 定位，[data-collapsed] 是 display: none——所以「不渲染」和「渲染后折叠」在布局上等价，surface 解析完成时不会发生位移。新增的 e2e 在 main 上验证为红，对一帧级别的回归来说这是最关键的一点。
Approve。行内一条 P3、一条说明，都不阻塞。
| bottomOpen: boolean; | ||
| }) { | ||
| const copy = getShellCopy(useUiLocale()).app; | ||
| if (props.hidden || (props.rightCollapsed && !props.bottomOpen)) return null; |
There was a problem hiding this comment.
[P3] This re-derives a rule that already has an owner. workbar-surface.tsx:742 computes the same fact as
constvisible=!props.hidden&&(placement==='right' ? !props.rightCollapsed : props.bottomOpen);and the fallback now expresses it a second time, in a different shape. The two agree today, which is why this is P3 — but a fallback drifting from the resolved surface is exactly the bug this PR is fixing, so leaving a second copy of the rule behind reopens the same seam.
The surface also renders both cards always and hides the invisible one with data-collapsed, while the fallback omits it instead. Same pixels (the CSS makes them equivalent), different DOM. Exporting one predicate and letting the fallback mirror the surface's structure collapses both differences, and the hidden || (rightCollapsed && !bottomOpen) early return then falls out — it is already implied by the placement list being empty.
| watch.visibleRightWorkbar = true; | ||
| } | ||
| }; | ||
| const observer = new MutationObserver(inspect); |
There was a problem hiding this comment.
The observer detects by sampling rather than by record: inspect() re-queries the live DOM, so a card added and removed inside one mutation batch leaves records behind but nothing for the query to find. It caught the real regression — you verified it red on main — so this is not a problem today, just the part that would quietly stop catching things. Reading the added nodes out of records would make the detection independent of how fast the flash is.
Uh oh!
There was an error while loading. Please reload this page.
Summary
Verification
origin/main: the new Electron test observedvisibleRightWorkbar === trueduring the first send.npm --workspace @maka/ui run buildnpx tsc -p apps/desktop/tsconfig.renderer.json --noEmit --pretty falsenpx biome lint apps/desktop/src/renderer/features/workbar/ui/workbar-host.tsx apps/desktop/e2e/session-workbar.spec.tsnpm --workspace @maka/desktop run build:renderernode --test apps/desktop/dist/main/__tests__/workbar-boundary.test.js apps/desktop/dist/main/__tests__/workbar-model.test.js(14/14 passed)Root cause
Creating the first session mounts
WorkbarHostwhile the lazyWorkbarSurfaceis still resolving. The old Suspense fallback always rendered a visible 480 px right panel, even whenrightCollapsedwas true. Once the real surface loaded, its collapsed state hid that panel, producing the brief flash. The fallback now derives its rendered placements from the samehidden,rightCollapsed, andbottomOpenstate as the resolved surface.AI use
Select exactly one:
Tool(s) and scope: Codex investigated the render path, implemented the fix and regression test, and ran the verification described above.
Checklist
Does this PR entail a change in behavior?