test(desktop): poll animation landings instead of guessing their duration by UncertaintyDeterminesYou4ndMe · Pull Request #2452 · apache/maka

UncertaintyDeterminesYou4ndMe · 2026-08-07T16:00:46Z

Part of #2389 (desktop e2e workspace; Pi TUI landed as #2448, Runtime Host as #2450, the alignment audit follows separately).

What changed

Five waits guessed at durations the page can be asked about directly:

Spec	Was	Now
`prompt-rail` × 4 (indicator glide ×2, hover swell, hover settle)	`waitForTimeout(500-600)` then one-shot geometry read	`expect(...).toPass()` polling the asserted geometry itself (the repo's existing idiom, cf. `composer-skill-invocation.spec.ts`) — ends the moment the transition lands, and only a failing test pays the budget
`new-messages-indicator` pre-scroll settle	`waitForTimeout(250)`	`waitForFunction` on the observable layout fact: `scrollHeight` stable across two consecutive animation frames

Audited and preserved (each wait is the contract)

quote-companion's two 500ms holds — real-timer negative windows now documented with their derivation: they outlive the hook's SELECTION_SETTLE_MS (350ms) to prove a dismissed/composer selection stays hidden once the settle window has elapsed. The absence is the contract; there is no earlier observable.
The 90ms drag cadence — deliberately paces mouse moves inside the 350ms settle window.
titlebar-identity's 1s one-shot toast-absence read — its comment already explains why a retrying matcher would mask the bug.
bot-onboarding's 1.15s/1.3s waits — synchronized against fixture-scheduled provider results, not animation guesses.

Timing

The three touched suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden). The gain is load-tolerance, not wall time: the fixed 500/600ms reads lost races on loaded CI runners (the #2221/#2332 failure shape), while the polls track the asserted fact under any scheduler. 15/15 tests pass before and after.

…tion Part of apache#2389. Four desktop e2e waits guessed at CSS transition durations with flat 500-600ms sleeps, and one guessed at layout settling with 250ms: - prompt-rail: the indicator glide, the rail's self-scroll alignment, and the hover swell/settle now poll the asserted geometry itself via the repo's existing expect(...).toPass() idiom — the wait ends the moment the transition lands and only a failing test pays the budget. - new-messages-indicator: the pre-scroll settle now waits for the observable layout fact (scrollHeight stable across two consecutive animation frames) before driving the extra scroll cycle. Preserved real-timer windows, now documented with their derivation: - quote-companion's two 500ms holds outlive the hook's SELECTION_SETTLE_MS (350ms) to prove the layer stays hidden after the settle window elapses — the absence is the contract, so the wait is the mechanism. - The 90ms drag cadence (deliberately inside the 350ms settle window), titlebar-identity's one-shot toast-absence read, and bot-onboarding's fixture-scheduled provider timing are untouched: each wait is the contract, not a guess. Timing: the three suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden); the change removes their load sensitivity — the polls track the asserted fact under any scheduler, where the fixed sleeps lost races on loaded CI runners. 15/15 tests pass.

UncertaintyDeterminesYou4ndMe · 2026-08-07T16:00:46Z

Part of #2389 (desktop e2e workspace; Pi TUI landed as #2448, Runtime Host as #2450, the alignment audit follows separately).

What changed

Five waits guessed at durations the page can be asked about directly:

Spec	Was	Now
`prompt-rail` × 4 (indicator glide ×2, hover swell, hover settle)	`waitForTimeout(500-600)` then one-shot geometry read	`expect(...).toPass()` polling the asserted geometry itself (the repo's existing idiom, cf. `composer-skill-invocation.spec.ts`) — ends the moment the transition lands, and only a failing test pays the budget
`new-messages-indicator` pre-scroll settle	`waitForTimeout(250)`	`waitForFunction` on the observable layout fact: `scrollHeight` stable across two consecutive animation frames

Audited and preserved (each wait is the contract)

quote-companion's two 500ms holds — real-timer negative windows now documented with their derivation: they outlive the hook's SELECTION_SETTLE_MS (350ms) to prove a dismissed/composer selection stays hidden once the settle window has elapsed. The absence is the contract; there is no earlier observable.
The 90ms drag cadence — deliberately paces mouse moves inside the 350ms settle window.
titlebar-identity's 1s one-shot toast-absence read — its comment already explains why a retrying matcher would mask the bug.
bot-onboarding's 1.15s/1.3s waits — synchronized against fixture-scheduled provider results, not animation guesses.

Timing

The three touched suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden). The gain is load-tolerance, not wall time: the fixed 500/600ms reads lost races on loaded CI runners (the #2221/#2332 failure shape), while the polls track the asserted fact under any scheduler. 15/15 tests pass before and after.

…tion Part of apache#2389. Four desktop e2e waits guessed at CSS transition durations with flat 500-600ms sleeps, and one guessed at layout settling with 250ms: - prompt-rail: the indicator glide, the rail's self-scroll alignment, and the hover swell/settle now poll the asserted geometry itself via the repo's existing expect(...).toPass() idiom — the wait ends the moment the transition lands and only a failing test pays the budget. - new-messages-indicator: the pre-scroll settle now waits for the observable layout fact (scrollHeight stable across two consecutive animation frames) before driving the extra scroll cycle. Preserved real-timer windows, now documented with their derivation: - quote-companion's two 500ms holds outlive the hook's SELECTION_SETTLE_MS (350ms) to prove the layer stays hidden after the settle window elapses — the absence is the contract, so the wait is the mechanism. - The 90ms drag cadence (deliberately inside the 350ms settle window), titlebar-identity's one-shot toast-absence read, and bot-onboarding's fixture-scheduled provider timing are untouched: each wait is the contract, not a guess. Timing: the three suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden); the change removes their load sensitivity — the polls track the asserted fact under any scheduler, where the fixed sleeps lost races on loaded CI runners. 15/15 tests pass.

UncertaintyDeterminesYou4ndMe · 2026-08-07T16:00:46Z

Part of #2389 (desktop e2e workspace; Pi TUI landed as #2448, Runtime Host as #2450, the alignment audit follows separately).

What changed

Five waits guessed at durations the page can be asked about directly:

Spec	Was	Now
`prompt-rail` × 4 (indicator glide ×2, hover swell, hover settle)	`waitForTimeout(500-600)` then one-shot geometry read	`expect(...).toPass()` polling the asserted geometry itself (the repo's existing idiom, cf. `composer-skill-invocation.spec.ts`) — ends the moment the transition lands, and only a failing test pays the budget
`new-messages-indicator` pre-scroll settle	`waitForTimeout(250)`	`waitForFunction` on the observable layout fact: `scrollHeight` stable across two consecutive animation frames

Audited and preserved (each wait is the contract)

quote-companion's two 500ms holds — real-timer negative windows now documented with their derivation: they outlive the hook's SELECTION_SETTLE_MS (350ms) to prove a dismissed/composer selection stays hidden once the settle window has elapsed. The absence is the contract; there is no earlier observable.
The 90ms drag cadence — deliberately paces mouse moves inside the 350ms settle window.
titlebar-identity's 1s one-shot toast-absence read — its comment already explains why a retrying matcher would mask the bug.
bot-onboarding's 1.15s/1.3s waits — synchronized against fixture-scheduled provider results, not animation guesses.

Timing

The three touched suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden). The gain is load-tolerance, not wall time: the fixed 500/600ms reads lost races on loaded CI runners (the #2221/#2332 failure shape), while the polls track the asserted fact under any scheduler. 15/15 tests pass before and after.

…tion Part of apache#2389. Four desktop e2e waits guessed at CSS transition durations with flat 500-600ms sleeps, and one guessed at layout settling with 250ms: - prompt-rail: the indicator glide, the rail's self-scroll alignment, and the hover swell/settle now poll the asserted geometry itself via the repo's existing expect(...).toPass() idiom — the wait ends the moment the transition lands and only a failing test pays the budget. - new-messages-indicator: the pre-scroll settle now waits for the observable layout fact (scrollHeight stable across two consecutive animation frames) before driving the extra scroll cycle. Preserved real-timer windows, now documented with their derivation: - quote-companion's two 500ms holds outlive the hook's SELECTION_SETTLE_MS (350ms) to prove the layer stays hidden after the settle window elapses — the absence is the contract, so the wait is the mechanism. - The 90ms drag cadence (deliberately inside the 350ms settle window), titlebar-identity's one-shot toast-absence read, and bot-onboarding's fixture-scheduled provider timing are untouched: each wait is the contract, not a guess. Timing: the three suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden); the change removes their load sensitivity — the polls track the asserted fact under any scheduler, where the fixed sleeps lost races on loaded CI runners. 15/15 tests pass.

UncertaintyDeterminesYou4ndMe · 2026-08-07T16:00:46Z

Part of #2389 (desktop e2e workspace; Pi TUI landed as #2448, Runtime Host as #2450, the alignment audit follows separately).

What changed

Five waits guessed at durations the page can be asked about directly:

Spec	Was	Now
`prompt-rail` × 4 (indicator glide ×2, hover swell, hover settle)	`waitForTimeout(500-600)` then one-shot geometry read	`expect(...).toPass()` polling the asserted geometry itself (the repo's existing idiom, cf. `composer-skill-invocation.spec.ts`) — ends the moment the transition lands, and only a failing test pays the budget
`new-messages-indicator` pre-scroll settle	`waitForTimeout(250)`	`waitForFunction` on the observable layout fact: `scrollHeight` stable across two consecutive animation frames

Audited and preserved (each wait is the contract)

quote-companion's two 500ms holds — real-timer negative windows now documented with their derivation: they outlive the hook's SELECTION_SETTLE_MS (350ms) to prove a dismissed/composer selection stays hidden once the settle window has elapsed. The absence is the contract; there is no earlier observable.
The 90ms drag cadence — deliberately paces mouse moves inside the 350ms settle window.
titlebar-identity's 1s one-shot toast-absence read — its comment already explains why a retrying matcher would mask the bug.
bot-onboarding's 1.15s/1.3s waits — synchronized against fixture-scheduled provider results, not animation guesses.

Timing

The three touched suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden). The gain is load-tolerance, not wall time: the fixed 500/600ms reads lost races on loaded CI runners (the #2221/#2332 failure shape), while the polls track the asserted fact under any scheduler. 15/15 tests pass before and after.

…tion Part of apache#2389. Four desktop e2e waits guessed at CSS transition durations with flat 500-600ms sleeps, and one guessed at layout settling with 250ms: - prompt-rail: the indicator glide, the rail's self-scroll alignment, and the hover swell/settle now poll the asserted geometry itself via the repo's existing expect(...).toPass() idiom — the wait ends the moment the transition lands and only a failing test pays the budget. - new-messages-indicator: the pre-scroll settle now waits for the observable layout fact (scrollHeight stable across two consecutive animation frames) before driving the extra scroll cycle. Preserved real-timer windows, now documented with their derivation: - quote-companion's two 500ms holds outlive the hook's SELECTION_SETTLE_MS (350ms) to prove the layer stays hidden after the settle window elapses — the absence is the contract, so the wait is the mechanism. - The 90ms drag cadence (deliberately inside the 350ms settle window), titlebar-identity's one-shot toast-absence read, and bot-onboarding's fixture-scheduled provider timing are untouched: each wait is the contract, not a guess. Timing: the three suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden); the change removes their load sensitivity — the polls track the asserted fact under any scheduler, where the fixed sleeps lost races on loaded CI runners. 15/15 tests pass.

UncertaintyDeterminesYou4ndMe · 2026-08-07T16:00:46Z

Part of #2389 (desktop e2e workspace; Pi TUI landed as #2448, Runtime Host as #2450, the alignment audit follows separately).

What changed

Five waits guessed at durations the page can be asked about directly:

Spec	Was	Now
`prompt-rail` × 4 (indicator glide ×2, hover swell, hover settle)	`waitForTimeout(500-600)` then one-shot geometry read	`expect(...).toPass()` polling the asserted geometry itself (the repo's existing idiom, cf. `composer-skill-invocation.spec.ts`) — ends the moment the transition lands, and only a failing test pays the budget
`new-messages-indicator` pre-scroll settle	`waitForTimeout(250)`	`waitForFunction` on the observable layout fact: `scrollHeight` stable across two consecutive animation frames

Audited and preserved (each wait is the contract)

quote-companion's two 500ms holds — real-timer negative windows now documented with their derivation: they outlive the hook's SELECTION_SETTLE_MS (350ms) to prove a dismissed/composer selection stays hidden once the settle window has elapsed. The absence is the contract; there is no earlier observable.
The 90ms drag cadence — deliberately paces mouse moves inside the 350ms settle window.
titlebar-identity's 1s one-shot toast-absence read — its comment already explains why a retrying matcher would mask the bug.
bot-onboarding's 1.15s/1.3s waits — synchronized against fixture-scheduled provider results, not animation guesses.

Timing

The three touched suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden). The gain is load-tolerance, not wall time: the fixed 500/600ms reads lost races on loaded CI runners (the #2221/#2332 failure shape), while the polls track the asserted fact under any scheduler. 15/15 tests pass before and after.

…tion Part of apache#2389. Four desktop e2e waits guessed at CSS transition durations with flat 500-600ms sleeps, and one guessed at layout settling with 250ms: - prompt-rail: the indicator glide, the rail's self-scroll alignment, and the hover swell/settle now poll the asserted geometry itself via the repo's existing expect(...).toPass() idiom — the wait ends the moment the transition lands and only a failing test pays the budget. - new-messages-indicator: the pre-scroll settle now waits for the observable layout fact (scrollHeight stable across two consecutive animation frames) before driving the extra scroll cycle. Preserved real-timer windows, now documented with their derivation: - quote-companion's two 500ms holds outlive the hook's SELECTION_SETTLE_MS (350ms) to prove the layer stays hidden after the settle window elapses — the absence is the contract, so the wait is the mechanism. - The 90ms drag cadence (deliberately inside the 350ms settle window), titlebar-identity's one-shot toast-absence read, and bot-onboarding's fixture-scheduled provider timing are untouched: each wait is the contract, not a guess. Timing: the three suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden); the change removes their load sensitivity — the polls track the asserted fact under any scheduler, where the fixed sleeps lost races on loaded CI runners. 15/15 tests pass.

UncertaintyDeterminesYou4ndMe · 2026-08-07T16:00:46Z

Part of #2389 (desktop e2e workspace; Pi TUI landed as #2448, Runtime Host as #2450, the alignment audit follows separately).

What changed

Five waits guessed at durations the page can be asked about directly:

Spec	Was	Now
`prompt-rail` × 4 (indicator glide ×2, hover swell, hover settle)	`waitForTimeout(500-600)` then one-shot geometry read	`expect(...).toPass()` polling the asserted geometry itself (the repo's existing idiom, cf. `composer-skill-invocation.spec.ts`) — ends the moment the transition lands, and only a failing test pays the budget
`new-messages-indicator` pre-scroll settle	`waitForTimeout(250)`	`waitForFunction` on the observable layout fact: `scrollHeight` stable across two consecutive animation frames

Audited and preserved (each wait is the contract)

quote-companion's two 500ms holds — real-timer negative windows now documented with their derivation: they outlive the hook's SELECTION_SETTLE_MS (350ms) to prove a dismissed/composer selection stays hidden once the settle window has elapsed. The absence is the contract; there is no earlier observable.
The 90ms drag cadence — deliberately paces mouse moves inside the 350ms settle window.
titlebar-identity's 1s one-shot toast-absence read — its comment already explains why a retrying matcher would mask the bug.
bot-onboarding's 1.15s/1.3s waits — synchronized against fixture-scheduled provider results, not animation guesses.

Timing

The three touched suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden). The gain is load-tolerance, not wall time: the fixed 500/600ms reads lost races on loaded CI runners (the #2221/#2332 failure shape), while the polls track the asserted fact under any scheduler. 15/15 tests pass before and after.

…tion Part of apache#2389. Four desktop e2e waits guessed at CSS transition durations with flat 500-600ms sleeps, and one guessed at layout settling with 250ms: - prompt-rail: the indicator glide, the rail's self-scroll alignment, and the hover swell/settle now poll the asserted geometry itself via the repo's existing expect(...).toPass() idiom — the wait ends the moment the transition lands and only a failing test pays the budget. - new-messages-indicator: the pre-scroll settle now waits for the observable layout fact (scrollHeight stable across two consecutive animation frames) before driving the extra scroll cycle. Preserved real-timer windows, now documented with their derivation: - quote-companion's two 500ms holds outlive the hook's SELECTION_SETTLE_MS (350ms) to prove the layer stays hidden after the settle window elapses — the absence is the contract, so the wait is the mechanism. - The 90ms drag cadence (deliberately inside the 350ms settle window), titlebar-identity's one-shot toast-absence read, and bot-onboarding's fixture-scheduled provider timing are untouched: each wait is the contract, not a guess. Timing: the three suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden); the change removes their load sensitivity — the polls track the asserted fact under any scheduler, where the fixed sleeps lost races on loaded CI runners. 15/15 tests pass.

UncertaintyDeterminesYou4ndMe · 2026-08-07T16:00:46Z

Part of #2389 (desktop e2e workspace; Pi TUI landed as #2448, Runtime Host as #2450, the alignment audit follows separately).

What changed

Five waits guessed at durations the page can be asked about directly:

Spec	Was	Now
`prompt-rail` × 4 (indicator glide ×2, hover swell, hover settle)	`waitForTimeout(500-600)` then one-shot geometry read	`expect(...).toPass()` polling the asserted geometry itself (the repo's existing idiom, cf. `composer-skill-invocation.spec.ts`) — ends the moment the transition lands, and only a failing test pays the budget
`new-messages-indicator` pre-scroll settle	`waitForTimeout(250)`	`waitForFunction` on the observable layout fact: `scrollHeight` stable across two consecutive animation frames

Audited and preserved (each wait is the contract)

quote-companion's two 500ms holds — real-timer negative windows now documented with their derivation: they outlive the hook's SELECTION_SETTLE_MS (350ms) to prove a dismissed/composer selection stays hidden once the settle window has elapsed. The absence is the contract; there is no earlier observable.
The 90ms drag cadence — deliberately paces mouse moves inside the 350ms settle window.
titlebar-identity's 1s one-shot toast-absence read — its comment already explains why a retrying matcher would mask the bug.
bot-onboarding's 1.15s/1.3s waits — synchronized against fixture-scheduled provider results, not animation guesses.

Timing

The three touched suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden). The gain is load-tolerance, not wall time: the fixed 500/600ms reads lost races on loaded CI runners (the #2221/#2332 failure shape), while the polls track the asserted fact under any scheduler. 15/15 tests pass before and after.

…tion Part of apache#2389. Four desktop e2e waits guessed at CSS transition durations with flat 500-600ms sleeps, and one guessed at layout settling with 250ms: - prompt-rail: the indicator glide, the rail's self-scroll alignment, and the hover swell/settle now poll the asserted geometry itself via the repo's existing expect(...).toPass() idiom — the wait ends the moment the transition lands and only a failing test pays the budget. - new-messages-indicator: the pre-scroll settle now waits for the observable layout fact (scrollHeight stable across two consecutive animation frames) before driving the extra scroll cycle. Preserved real-timer windows, now documented with their derivation: - quote-companion's two 500ms holds outlive the hook's SELECTION_SETTLE_MS (350ms) to prove the layer stays hidden after the settle window elapses — the absence is the contract, so the wait is the mechanism. - The 90ms drag cadence (deliberately inside the 350ms settle window), titlebar-identity's one-shot toast-absence read, and bot-onboarding's fixture-scheduled provider timing are untouched: each wait is the contract, not a guess. Timing: the three suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden); the change removes their load sensitivity — the polls track the asserted fact under any scheduler, where the fixed sleeps lost races on loaded CI runners. 15/15 tests pass.

UncertaintyDeterminesYou4ndMe · 2026-08-07T16:00:46Z

Part of #2389 (desktop e2e workspace; Pi TUI landed as #2448, Runtime Host as #2450, the alignment audit follows separately).

What changed

Five waits guessed at durations the page can be asked about directly:

Spec	Was	Now
`prompt-rail` × 4 (indicator glide ×2, hover swell, hover settle)	`waitForTimeout(500-600)` then one-shot geometry read	`expect(...).toPass()` polling the asserted geometry itself (the repo's existing idiom, cf. `composer-skill-invocation.spec.ts`) — ends the moment the transition lands, and only a failing test pays the budget
`new-messages-indicator` pre-scroll settle	`waitForTimeout(250)`	`waitForFunction` on the observable layout fact: `scrollHeight` stable across two consecutive animation frames

Audited and preserved (each wait is the contract)

quote-companion's two 500ms holds — real-timer negative windows now documented with their derivation: they outlive the hook's SELECTION_SETTLE_MS (350ms) to prove a dismissed/composer selection stays hidden once the settle window has elapsed. The absence is the contract; there is no earlier observable.
The 90ms drag cadence — deliberately paces mouse moves inside the 350ms settle window.
titlebar-identity's 1s one-shot toast-absence read — its comment already explains why a retrying matcher would mask the bug.
bot-onboarding's 1.15s/1.3s waits — synchronized against fixture-scheduled provider results, not animation guesses.

Timing

The three touched suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden). The gain is load-tolerance, not wall time: the fixed 500/600ms reads lost races on loaded CI runners (the #2221/#2332 failure shape), while the polls track the asserted fact under any scheduler. 15/15 tests pass before and after.

…tion Part of apache#2389. Four desktop e2e waits guessed at CSS transition durations with flat 500-600ms sleeps, and one guessed at layout settling with 250ms: - prompt-rail: the indicator glide, the rail's self-scroll alignment, and the hover swell/settle now poll the asserted geometry itself via the repo's existing expect(...).toPass() idiom — the wait ends the moment the transition lands and only a failing test pays the budget. - new-messages-indicator: the pre-scroll settle now waits for the observable layout fact (scrollHeight stable across two consecutive animation frames) before driving the extra scroll cycle. Preserved real-timer windows, now documented with their derivation: - quote-companion's two 500ms holds outlive the hook's SELECTION_SETTLE_MS (350ms) to prove the layer stays hidden after the settle window elapses — the absence is the contract, so the wait is the mechanism. - The 90ms drag cadence (deliberately inside the 350ms settle window), titlebar-identity's one-shot toast-absence read, and bot-onboarding's fixture-scheduled provider timing are untouched: each wait is the contract, not a guess. Timing: the three suites run in the same ~20s wall locally (Playwright overlaps workers, so the removed sleeps were latency-hidden); the change removes their load sensitivity — the polls track the asserted fact under any scheduler, where the fixed sleeps lost races on loaded CI runners. 15/15 tests pass.

UncertaintyDeterminesYou4ndMe mentioned this pull request Aug 7, 2026
perf(test): replace fixed waits with observable completion #2389
Closed

jackwener merged commit 717c6c5 into apache:mainAug 7, 2026
12 checks passed

jackwener mentioned this pull request Aug 8, 2026
test(desktop): merge same-fixture e2e assertions into coherent journeys #2489
Merged