flaky(e2e): transcript-scroll:498 asserts the tail before a viewport change re-pins it

[Astro-Han]

What happened

The Desktop E2E test a nested scroller near the history boundary does not request an earlier range (apps/desktop/e2e/transcript-scroll.spec.ts:498) fails intermittently on CI, before it exercises anything it is about.

The test resizes the viewport, waits six painted frames, and then asserts its two preconditions in one shot:

await page.setViewportSize({ width: 900, height: 1500 });
await waitForPaintedFrames(page, 6);
const metrics = await scrollMetrics(page);
expect(metrics.scrollTop).toBeLessThanOrEqual(Math.max(640, metrics.clientHeight * 2));
expect(metrics.distance).toBeLessThanOrEqual(4);   // <- fails here

Observed on CI:

Error: expect(received).toBeLessThanOrEqual(expected)
Expected: <= 4
Received:    58
  at apps/desktop/e2e/transcript-scroll.spec.ts:505:28

A viewport change makes the transcript re-pin to the tail. Six painted frames is a guess at how long that takes, not a wait for it, so under CI load the assertion samples the scroller mid-pin and reads a distance of 58px. Nothing about the nested-scroller behaviour under test has run yet.

This is the same failure mode as #4619 (fixed in #4620) — asserting on transcript geometry before tail positioning settles — in a different test in the same file.

How to reproduce

Intermittent; it does not reproduce reliably on a developer machine (11/11 green locally, that test in 528ms). It appears under concurrent CI load, where the Desktop E2E job runs 109 tests across 4 workers on Xvfb.

1. Push any branch and let the Desktop e2e job run.
2. Observed twice today: apache/maka on main at 148f8eb (run 33737060147) and on PR refactor(runtime): converge AgentRun metadata into the RuntimeInvocation event spine #4631 (run 33741800324), both with Expected: <= 4 / Received: 58.

Rate over the last 20 CI runs on main: 13 completed (7 cancelled by newer pushes), 3 red, all in the scroll family — code-scroll.spec.ts:22 twice (fixed since, in #4632), this test once, prompt-rail.spec.ts:370 once. So this one is roughly 1 in 13 runs, and the job as a whole about 1 in 4.

Environment

• Maka commit: main at 148f8eb, also seen on PR refactor(runtime): converge AgentRun metadata into the RuntimeInvocation event spine #4631
• OS and version: CI runner, Linux + Xvfb
• Surface: Desktop
• Node.js version, if running from source: as pinned by the release npm toolchain in CI

Logs, screenshots, or additional context

Suggested fix. Replace the one-shot sample with a poll for the state the test needs, the way #4620 did for its neighbour a few lines below. The precondition is two facts at once — the reader sits near the history boundary and the transcript is pinned to the tail — so poll them together rather than asserting them separately after a fixed number of frames:

await page.setViewportSize({ width: 900, height: 1500 });
await expect.poll(async () => {
  const metrics = await scrollMetrics(page);
  return {
    nearHistoryBoundary: metrics.scrollTop <= Math.max(640, metrics.clientHeight * 2),
    settledAtTail: metrics.distance <= 4,
  };
}, { message: 'the transcript re-pins to the tail after the viewport change' })
  .toMatchObject({ nearHistoryBoundary: true, settledAtTail: true });

Two things worth keeping in mind while fixing it:

• Do not relax the <= 4 tolerance. The number is the point of the test: the whole file exists to prove the transcript stays under the reader, and a looser bound would let a real release of the tail pass.
• The same one-shot pattern appears elsewhere in this file after actions that move the scroller. Converting only this call site is a complete fix for this issue; converting the others is welcome but belongs in its own commit, and each conversion should say which action it is waiting out.

Verify with npm exec -w @maka/desktop -- playwright test --config e2e/playwright.config.ts e2e/transcript-scroll.spec.ts. Because the failure is load-dependent, a useful check is to run that file repeatedly with --repeat-each and several workers rather than once.

This is a self-contained change in one test file with a worked example in the tree, so it is a reasonable entry point for anyone who already has the Desktop E2E setup running. To claim it, comment take (see CONTRIBUTING.md).

[CxHsin]

take

[CxHsin]

This issue was filed before the fix landed. It was subsequently fixed by commit 4149d68 in PR #4386, merged into apache:main on Sep 3, 2026. The target test now polls the transcript geometry after the viewport change instead of taking a one-shot sample.