test(desktop): merge same-fixture e2e assertions into coherent journeys by UncertaintyDeterminesYou4ndMe · Pull Request #2486 · apache/maka

UncertaintyDeterminesYou4ndMe · 2026-08-08T07:21:11Z

Part of #2390 (Electron E2E half; the Storybook matrix reduction is #2481 per the issue's separate-PR requirement).

Problem

Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — and most of them re-launched and re-seeded exactly the state their file-mates had just built. quote-companion.spec.ts alone launched 12 windows over one identical seed; new-messages-indicator.spec.ts re-ran its six-message transcript seed from scratch for its second test.

Change

Tests that share one fixture and compatible assertions now run as consecutive phases of one journey. 69 tests / 69 launches → 34 tests / 34 launches. Every merge is annotated in-file with the ordering constraint that made it safe:

Count pins run first (e.g. 总消息数 === 1 phases, the blocked-Skill phase that pins zero turns and zero sessions).
Destructive phases run last (rename in settings-projects, the don't-ask-again checkbox that suppresses close confirmations for the rest of a window).
Turn-settle waits guard every phase boundary that sends — an Enter during a streaming turn silently becomes steering, which two merges surfaced immediately.

file	tests	file	tests
quote-companion	12 → 4	composer-skill-invocation	7 → 3
settings	6 → 2	mcp	5 → 2
composer-mention-token	4 → 1	send-message	4 → 2
providers	3 → 1	settings-projects	3 → 1
7 two-test files	2 → 1 each

Also removes 11 window fixtures no spec references (dead seeding code left behind by earlier test deletions, fixtures.ts 592 → 427 lines).

Deliberately not merged, and why

storage-root-conflict: the cold-start race is the subject; sharing a window would erase the precondition.
permission-mode-surface: its two tests boot two different fixture scenarios.
The staged-Skills draft-restoration test keeps its own window — observed, not guessed: merged into a shared window, its final send serialized raw /skill: text instead of chips, because concurrent session activity triggers exactly the external-value-change editor rebuild the test guards. That failure mode is worth a launch.
quote-companion's batch-close: needs the confirmation dialog the numbered-tab journey suppresses.
Cross-file candidates (the invocableSkillsWindow family spans 5 files; sandboxBoundaryWindow spans 2) are left as a maintainer decision — merging across files moves regression ownership between files, which is an organizational call, not a mechanical one.

Distinct contracts preserved

All IPC, focus, scrolling, remount, persistence, lifecycle, and animation assertions are unchanged — phases are the original test bodies with their assertions intact (three assertions were strengthened in passing: hostile-Mermaid now pins settled-turn count 2, the markdown --markdown scan asserts exit 0 explicitly, ask-user-question now answers after the reload, proving the rehydrated prompt is the live parked turn).

Timing

Same machine, warm build; CI runs workers=1 where launch count dominates:

config	before	after
workers=4 (local)	69 passed, 57.9s	34 passed, 47.3 / 43.2 / 40.8s (3 rounds)
workers=1 (CI-like)	69 passed, 192.4s	34 passed, 147.2s (−23%)

Electron launches: 69 → 34. Typecheck and biome are clean.

cc @Astro-Han

Part of apache#2390 (Electron E2E half; the Storybook matrix is a separate PR per the issue). Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — with most tests re-launching and re-seeding the exact state their file-mates had just built. Tests that share one fixture and compatible assertions now run as consecutive phases of one journey; every merge is annotated with the ordering constraint that made it safe (count pins run first, destructive phases run last, don't-ask-again phases end their window). Merged: ask-user-question 2->1, attachment 2->1, bot-onboarding 2->1, composer-mention-token 4->1, composer-skill-invocation 7->3, keyboard-help 2->1, mcp 5->2, new-messages-indicator 2->1 (also stops re-seeding the six-message transcript), providers 3->1, quote-companion 12->4, send-message 4->2, session-workbar 2->1, settings-projects 3->1, settings 6->2, skill-delete-scope 2->1. Kept separate deliberately: - storage-root-conflict (cold-start race is the subject), - permission-mode-surface (two different fixture scenarios), - the staged-Skills draft-restoration test (its contract is editor rebuilds on external value changes; concurrent session activity in a shared window perturbs exactly that — observed, not guessed), - quote-companion's batch-close (needs the confirmation dialog the numbered-tab journey suppresses via don't-ask-again). Also removes 11 window fixtures that no spec references (dead seeding code left behind by earlier test deletions): longTranscriptWindow, shortFinalTurnWindow, overflowingRailWindow, sidebarLongSessionsWindow, disclosureOutputWindow, staleSessionsWindow, gitReviewWindow, gitReviewLargeWindow, artifactPaneWindow, localeSwitchWindow, planRemindersWindow. 69 tests / 69 launches -> 34 tests / 34 launches. Local (4 workers): 57.9s -> 47.3/43.2/40.8s across three green rounds. CI runs a single worker, where launch count dominates wall time.

jackwener · 2026-08-08T07:29:57Z

Review decision

Problem is real: function-scoped fixtures re-launch Electron for identical seeds; launch count dominates CI wall time.

Approach is correct: merge same-fixture assertions into ordered phases (count pins first, destructive last, turn-settle between sends), keep must-isolate cases (storage-root-conflict, permission-mode dual scenarios, staged-Skills draft restore).

Landing via maintainer branch because the fork head conflicts with main after #2478 presentation cleanup. Rebase drops reintroduced presentation journeys (MCP layout matrix, providers header geometry, settings back-icon rail) while keeping product journeys and dead-fixture removal.

See the rebased PR for the final tree.

jackwener · 2026-08-08T07:30:08Z

Landing via rebased #2489 (presentation pins already removed on main via #2478 are not reintroduced). Thanks @UncertaintyDeterminesYou4ndMe!

UncertaintyDeterminesYou4ndMe · 2026-08-08T07:21:11Z

Part of #2390 (Electron E2E half; the Storybook matrix reduction is #2481 per the issue's separate-PR requirement).

Problem

Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — and most of them re-launched and re-seeded exactly the state their file-mates had just built. quote-companion.spec.ts alone launched 12 windows over one identical seed; new-messages-indicator.spec.ts re-ran its six-message transcript seed from scratch for its second test.

Change

Tests that share one fixture and compatible assertions now run as consecutive phases of one journey. 69 tests / 69 launches → 34 tests / 34 launches. Every merge is annotated in-file with the ordering constraint that made it safe:

Count pins run first (e.g. 总消息数 === 1 phases, the blocked-Skill phase that pins zero turns and zero sessions).
Destructive phases run last (rename in settings-projects, the don't-ask-again checkbox that suppresses close confirmations for the rest of a window).
Turn-settle waits guard every phase boundary that sends — an Enter during a streaming turn silently becomes steering, which two merges surfaced immediately.

file	tests	file	tests
quote-companion	12 → 4	composer-skill-invocation	7 → 3
settings	6 → 2	mcp	5 → 2
composer-mention-token	4 → 1	send-message	4 → 2
providers	3 → 1	settings-projects	3 → 1
7 two-test files	2 → 1 each

Also removes 11 window fixtures no spec references (dead seeding code left behind by earlier test deletions, fixtures.ts 592 → 427 lines).

Deliberately not merged, and why

storage-root-conflict: the cold-start race is the subject; sharing a window would erase the precondition.
permission-mode-surface: its two tests boot two different fixture scenarios.
The staged-Skills draft-restoration test keeps its own window — observed, not guessed: merged into a shared window, its final send serialized raw /skill: text instead of chips, because concurrent session activity triggers exactly the external-value-change editor rebuild the test guards. That failure mode is worth a launch.
quote-companion's batch-close: needs the confirmation dialog the numbered-tab journey suppresses.
Cross-file candidates (the invocableSkillsWindow family spans 5 files; sandboxBoundaryWindow spans 2) are left as a maintainer decision — merging across files moves regression ownership between files, which is an organizational call, not a mechanical one.

Distinct contracts preserved

All IPC, focus, scrolling, remount, persistence, lifecycle, and animation assertions are unchanged — phases are the original test bodies with their assertions intact (three assertions were strengthened in passing: hostile-Mermaid now pins settled-turn count 2, the markdown --markdown scan asserts exit 0 explicitly, ask-user-question now answers after the reload, proving the rehydrated prompt is the live parked turn).

Timing

Same machine, warm build; CI runs workers=1 where launch count dominates:

config	before	after
workers=4 (local)	69 passed, 57.9s	34 passed, 47.3 / 43.2 / 40.8s (3 rounds)
workers=1 (CI-like)	69 passed, 192.4s	34 passed, 147.2s (−23%)

Electron launches: 69 → 34. Typecheck and biome are clean.

cc @Astro-Han

Part of apache#2390 (Electron E2E half; the Storybook matrix is a separate PR per the issue). Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — with most tests re-launching and re-seeding the exact state their file-mates had just built. Tests that share one fixture and compatible assertions now run as consecutive phases of one journey; every merge is annotated with the ordering constraint that made it safe (count pins run first, destructive phases run last, don't-ask-again phases end their window). Merged: ask-user-question 2->1, attachment 2->1, bot-onboarding 2->1, composer-mention-token 4->1, composer-skill-invocation 7->3, keyboard-help 2->1, mcp 5->2, new-messages-indicator 2->1 (also stops re-seeding the six-message transcript), providers 3->1, quote-companion 12->4, send-message 4->2, session-workbar 2->1, settings-projects 3->1, settings 6->2, skill-delete-scope 2->1. Kept separate deliberately: - storage-root-conflict (cold-start race is the subject), - permission-mode-surface (two different fixture scenarios), - the staged-Skills draft-restoration test (its contract is editor rebuilds on external value changes; concurrent session activity in a shared window perturbs exactly that — observed, not guessed), - quote-companion's batch-close (needs the confirmation dialog the numbered-tab journey suppresses via don't-ask-again). Also removes 11 window fixtures that no spec references (dead seeding code left behind by earlier test deletions): longTranscriptWindow, shortFinalTurnWindow, overflowingRailWindow, sidebarLongSessionsWindow, disclosureOutputWindow, staleSessionsWindow, gitReviewWindow, gitReviewLargeWindow, artifactPaneWindow, localeSwitchWindow, planRemindersWindow. 69 tests / 69 launches -> 34 tests / 34 launches. Local (4 workers): 57.9s -> 47.3/43.2/40.8s across three green rounds. CI runs a single worker, where launch count dominates wall time.

jackwener · 2026-08-08T07:29:57Z

Review decision

Problem is real: function-scoped fixtures re-launch Electron for identical seeds; launch count dominates CI wall time.

Approach is correct: merge same-fixture assertions into ordered phases (count pins first, destructive last, turn-settle between sends), keep must-isolate cases (storage-root-conflict, permission-mode dual scenarios, staged-Skills draft restore).

Landing via maintainer branch because the fork head conflicts with main after #2478 presentation cleanup. Rebase drops reintroduced presentation journeys (MCP layout matrix, providers header geometry, settings back-icon rail) while keeping product journeys and dead-fixture removal.

See the rebased PR for the final tree.

jackwener · 2026-08-08T07:30:08Z

Landing via rebased #2489 (presentation pins already removed on main via #2478 are not reintroduced). Thanks @UncertaintyDeterminesYou4ndMe!

UncertaintyDeterminesYou4ndMe · 2026-08-08T07:21:11Z

Part of #2390 (Electron E2E half; the Storybook matrix reduction is #2481 per the issue's separate-PR requirement).

Problem

Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — and most of them re-launched and re-seeded exactly the state their file-mates had just built. quote-companion.spec.ts alone launched 12 windows over one identical seed; new-messages-indicator.spec.ts re-ran its six-message transcript seed from scratch for its second test.

Change

Tests that share one fixture and compatible assertions now run as consecutive phases of one journey. 69 tests / 69 launches → 34 tests / 34 launches. Every merge is annotated in-file with the ordering constraint that made it safe:

Count pins run first (e.g. 总消息数 === 1 phases, the blocked-Skill phase that pins zero turns and zero sessions).
Destructive phases run last (rename in settings-projects, the don't-ask-again checkbox that suppresses close confirmations for the rest of a window).
Turn-settle waits guard every phase boundary that sends — an Enter during a streaming turn silently becomes steering, which two merges surfaced immediately.

file	tests	file	tests
quote-companion	12 → 4	composer-skill-invocation	7 → 3
settings	6 → 2	mcp	5 → 2
composer-mention-token	4 → 1	send-message	4 → 2
providers	3 → 1	settings-projects	3 → 1
7 two-test files	2 → 1 each

Also removes 11 window fixtures no spec references (dead seeding code left behind by earlier test deletions, fixtures.ts 592 → 427 lines).

Deliberately not merged, and why

storage-root-conflict: the cold-start race is the subject; sharing a window would erase the precondition.
permission-mode-surface: its two tests boot two different fixture scenarios.
The staged-Skills draft-restoration test keeps its own window — observed, not guessed: merged into a shared window, its final send serialized raw /skill: text instead of chips, because concurrent session activity triggers exactly the external-value-change editor rebuild the test guards. That failure mode is worth a launch.
quote-companion's batch-close: needs the confirmation dialog the numbered-tab journey suppresses.
Cross-file candidates (the invocableSkillsWindow family spans 5 files; sandboxBoundaryWindow spans 2) are left as a maintainer decision — merging across files moves regression ownership between files, which is an organizational call, not a mechanical one.

Distinct contracts preserved

All IPC, focus, scrolling, remount, persistence, lifecycle, and animation assertions are unchanged — phases are the original test bodies with their assertions intact (three assertions were strengthened in passing: hostile-Mermaid now pins settled-turn count 2, the markdown --markdown scan asserts exit 0 explicitly, ask-user-question now answers after the reload, proving the rehydrated prompt is the live parked turn).

Timing

Same machine, warm build; CI runs workers=1 where launch count dominates:

config	before	after
workers=4 (local)	69 passed, 57.9s	34 passed, 47.3 / 43.2 / 40.8s (3 rounds)
workers=1 (CI-like)	69 passed, 192.4s	34 passed, 147.2s (−23%)

Electron launches: 69 → 34. Typecheck and biome are clean.

cc @Astro-Han

Part of apache#2390 (Electron E2E half; the Storybook matrix is a separate PR per the issue). Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — with most tests re-launching and re-seeding the exact state their file-mates had just built. Tests that share one fixture and compatible assertions now run as consecutive phases of one journey; every merge is annotated with the ordering constraint that made it safe (count pins run first, destructive phases run last, don't-ask-again phases end their window). Merged: ask-user-question 2->1, attachment 2->1, bot-onboarding 2->1, composer-mention-token 4->1, composer-skill-invocation 7->3, keyboard-help 2->1, mcp 5->2, new-messages-indicator 2->1 (also stops re-seeding the six-message transcript), providers 3->1, quote-companion 12->4, send-message 4->2, session-workbar 2->1, settings-projects 3->1, settings 6->2, skill-delete-scope 2->1. Kept separate deliberately: - storage-root-conflict (cold-start race is the subject), - permission-mode-surface (two different fixture scenarios), - the staged-Skills draft-restoration test (its contract is editor rebuilds on external value changes; concurrent session activity in a shared window perturbs exactly that — observed, not guessed), - quote-companion's batch-close (needs the confirmation dialog the numbered-tab journey suppresses via don't-ask-again). Also removes 11 window fixtures that no spec references (dead seeding code left behind by earlier test deletions): longTranscriptWindow, shortFinalTurnWindow, overflowingRailWindow, sidebarLongSessionsWindow, disclosureOutputWindow, staleSessionsWindow, gitReviewWindow, gitReviewLargeWindow, artifactPaneWindow, localeSwitchWindow, planRemindersWindow. 69 tests / 69 launches -> 34 tests / 34 launches. Local (4 workers): 57.9s -> 47.3/43.2/40.8s across three green rounds. CI runs a single worker, where launch count dominates wall time.

jackwener · 2026-08-08T07:29:57Z

Review decision

Problem is real: function-scoped fixtures re-launch Electron for identical seeds; launch count dominates CI wall time.

Approach is correct: merge same-fixture assertions into ordered phases (count pins first, destructive last, turn-settle between sends), keep must-isolate cases (storage-root-conflict, permission-mode dual scenarios, staged-Skills draft restore).

Landing via maintainer branch because the fork head conflicts with main after #2478 presentation cleanup. Rebase drops reintroduced presentation journeys (MCP layout matrix, providers header geometry, settings back-icon rail) while keeping product journeys and dead-fixture removal.

See the rebased PR for the final tree.

jackwener · 2026-08-08T07:30:08Z

Landing via rebased #2489 (presentation pins already removed on main via #2478 are not reintroduced). Thanks @UncertaintyDeterminesYou4ndMe!

UncertaintyDeterminesYou4ndMe · 2026-08-08T07:21:11Z

Part of #2390 (Electron E2E half; the Storybook matrix reduction is #2481 per the issue's separate-PR requirement).

Problem

Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — and most of them re-launched and re-seeded exactly the state their file-mates had just built. quote-companion.spec.ts alone launched 12 windows over one identical seed; new-messages-indicator.spec.ts re-ran its six-message transcript seed from scratch for its second test.

Change

Tests that share one fixture and compatible assertions now run as consecutive phases of one journey. 69 tests / 69 launches → 34 tests / 34 launches. Every merge is annotated in-file with the ordering constraint that made it safe:

Count pins run first (e.g. 总消息数 === 1 phases, the blocked-Skill phase that pins zero turns and zero sessions).
Destructive phases run last (rename in settings-projects, the don't-ask-again checkbox that suppresses close confirmations for the rest of a window).
Turn-settle waits guard every phase boundary that sends — an Enter during a streaming turn silently becomes steering, which two merges surfaced immediately.

file	tests	file	tests
quote-companion	12 → 4	composer-skill-invocation	7 → 3
settings	6 → 2	mcp	5 → 2
composer-mention-token	4 → 1	send-message	4 → 2
providers	3 → 1	settings-projects	3 → 1
7 two-test files	2 → 1 each

Also removes 11 window fixtures no spec references (dead seeding code left behind by earlier test deletions, fixtures.ts 592 → 427 lines).

Deliberately not merged, and why

storage-root-conflict: the cold-start race is the subject; sharing a window would erase the precondition.
permission-mode-surface: its two tests boot two different fixture scenarios.
The staged-Skills draft-restoration test keeps its own window — observed, not guessed: merged into a shared window, its final send serialized raw /skill: text instead of chips, because concurrent session activity triggers exactly the external-value-change editor rebuild the test guards. That failure mode is worth a launch.
quote-companion's batch-close: needs the confirmation dialog the numbered-tab journey suppresses.
Cross-file candidates (the invocableSkillsWindow family spans 5 files; sandboxBoundaryWindow spans 2) are left as a maintainer decision — merging across files moves regression ownership between files, which is an organizational call, not a mechanical one.

Distinct contracts preserved

All IPC, focus, scrolling, remount, persistence, lifecycle, and animation assertions are unchanged — phases are the original test bodies with their assertions intact (three assertions were strengthened in passing: hostile-Mermaid now pins settled-turn count 2, the markdown --markdown scan asserts exit 0 explicitly, ask-user-question now answers after the reload, proving the rehydrated prompt is the live parked turn).

Timing

Same machine, warm build; CI runs workers=1 where launch count dominates:

config	before	after
workers=4 (local)	69 passed, 57.9s	34 passed, 47.3 / 43.2 / 40.8s (3 rounds)
workers=1 (CI-like)	69 passed, 192.4s	34 passed, 147.2s (−23%)

Electron launches: 69 → 34. Typecheck and biome are clean.

cc @Astro-Han

Part of apache#2390 (Electron E2E half; the Storybook matrix is a separate PR per the issue). Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — with most tests re-launching and re-seeding the exact state their file-mates had just built. Tests that share one fixture and compatible assertions now run as consecutive phases of one journey; every merge is annotated with the ordering constraint that made it safe (count pins run first, destructive phases run last, don't-ask-again phases end their window). Merged: ask-user-question 2->1, attachment 2->1, bot-onboarding 2->1, composer-mention-token 4->1, composer-skill-invocation 7->3, keyboard-help 2->1, mcp 5->2, new-messages-indicator 2->1 (also stops re-seeding the six-message transcript), providers 3->1, quote-companion 12->4, send-message 4->2, session-workbar 2->1, settings-projects 3->1, settings 6->2, skill-delete-scope 2->1. Kept separate deliberately: - storage-root-conflict (cold-start race is the subject), - permission-mode-surface (two different fixture scenarios), - the staged-Skills draft-restoration test (its contract is editor rebuilds on external value changes; concurrent session activity in a shared window perturbs exactly that — observed, not guessed), - quote-companion's batch-close (needs the confirmation dialog the numbered-tab journey suppresses via don't-ask-again). Also removes 11 window fixtures that no spec references (dead seeding code left behind by earlier test deletions): longTranscriptWindow, shortFinalTurnWindow, overflowingRailWindow, sidebarLongSessionsWindow, disclosureOutputWindow, staleSessionsWindow, gitReviewWindow, gitReviewLargeWindow, artifactPaneWindow, localeSwitchWindow, planRemindersWindow. 69 tests / 69 launches -> 34 tests / 34 launches. Local (4 workers): 57.9s -> 47.3/43.2/40.8s across three green rounds. CI runs a single worker, where launch count dominates wall time.

jackwener · 2026-08-08T07:29:57Z

Review decision

Problem is real: function-scoped fixtures re-launch Electron for identical seeds; launch count dominates CI wall time.

Approach is correct: merge same-fixture assertions into ordered phases (count pins first, destructive last, turn-settle between sends), keep must-isolate cases (storage-root-conflict, permission-mode dual scenarios, staged-Skills draft restore).

Landing via maintainer branch because the fork head conflicts with main after #2478 presentation cleanup. Rebase drops reintroduced presentation journeys (MCP layout matrix, providers header geometry, settings back-icon rail) while keeping product journeys and dead-fixture removal.

See the rebased PR for the final tree.

jackwener · 2026-08-08T07:30:08Z

Landing via rebased #2489 (presentation pins already removed on main via #2478 are not reintroduced). Thanks @UncertaintyDeterminesYou4ndMe!

UncertaintyDeterminesYou4ndMe · 2026-08-08T07:21:11Z

Part of #2390 (Electron E2E half; the Storybook matrix reduction is #2481 per the issue's separate-PR requirement).

Problem

Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — and most of them re-launched and re-seeded exactly the state their file-mates had just built. quote-companion.spec.ts alone launched 12 windows over one identical seed; new-messages-indicator.spec.ts re-ran its six-message transcript seed from scratch for its second test.

Change

Tests that share one fixture and compatible assertions now run as consecutive phases of one journey. 69 tests / 69 launches → 34 tests / 34 launches. Every merge is annotated in-file with the ordering constraint that made it safe:

Count pins run first (e.g. 总消息数 === 1 phases, the blocked-Skill phase that pins zero turns and zero sessions).
Destructive phases run last (rename in settings-projects, the don't-ask-again checkbox that suppresses close confirmations for the rest of a window).
Turn-settle waits guard every phase boundary that sends — an Enter during a streaming turn silently becomes steering, which two merges surfaced immediately.

file	tests	file	tests
quote-companion	12 → 4	composer-skill-invocation	7 → 3
settings	6 → 2	mcp	5 → 2
composer-mention-token	4 → 1	send-message	4 → 2
providers	3 → 1	settings-projects	3 → 1
7 two-test files	2 → 1 each

Also removes 11 window fixtures no spec references (dead seeding code left behind by earlier test deletions, fixtures.ts 592 → 427 lines).

Deliberately not merged, and why

storage-root-conflict: the cold-start race is the subject; sharing a window would erase the precondition.
permission-mode-surface: its two tests boot two different fixture scenarios.
The staged-Skills draft-restoration test keeps its own window — observed, not guessed: merged into a shared window, its final send serialized raw /skill: text instead of chips, because concurrent session activity triggers exactly the external-value-change editor rebuild the test guards. That failure mode is worth a launch.
quote-companion's batch-close: needs the confirmation dialog the numbered-tab journey suppresses.
Cross-file candidates (the invocableSkillsWindow family spans 5 files; sandboxBoundaryWindow spans 2) are left as a maintainer decision — merging across files moves regression ownership between files, which is an organizational call, not a mechanical one.

Distinct contracts preserved

All IPC, focus, scrolling, remount, persistence, lifecycle, and animation assertions are unchanged — phases are the original test bodies with their assertions intact (three assertions were strengthened in passing: hostile-Mermaid now pins settled-turn count 2, the markdown --markdown scan asserts exit 0 explicitly, ask-user-question now answers after the reload, proving the rehydrated prompt is the live parked turn).

Timing

Same machine, warm build; CI runs workers=1 where launch count dominates:

config	before	after
workers=4 (local)	69 passed, 57.9s	34 passed, 47.3 / 43.2 / 40.8s (3 rounds)
workers=1 (CI-like)	69 passed, 192.4s	34 passed, 147.2s (−23%)

Electron launches: 69 → 34. Typecheck and biome are clean.

cc @Astro-Han

Part of apache#2390 (Electron E2E half; the Storybook matrix is a separate PR per the issue). Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — with most tests re-launching and re-seeding the exact state their file-mates had just built. Tests that share one fixture and compatible assertions now run as consecutive phases of one journey; every merge is annotated with the ordering constraint that made it safe (count pins run first, destructive phases run last, don't-ask-again phases end their window). Merged: ask-user-question 2->1, attachment 2->1, bot-onboarding 2->1, composer-mention-token 4->1, composer-skill-invocation 7->3, keyboard-help 2->1, mcp 5->2, new-messages-indicator 2->1 (also stops re-seeding the six-message transcript), providers 3->1, quote-companion 12->4, send-message 4->2, session-workbar 2->1, settings-projects 3->1, settings 6->2, skill-delete-scope 2->1. Kept separate deliberately: - storage-root-conflict (cold-start race is the subject), - permission-mode-surface (two different fixture scenarios), - the staged-Skills draft-restoration test (its contract is editor rebuilds on external value changes; concurrent session activity in a shared window perturbs exactly that — observed, not guessed), - quote-companion's batch-close (needs the confirmation dialog the numbered-tab journey suppresses via don't-ask-again). Also removes 11 window fixtures that no spec references (dead seeding code left behind by earlier test deletions): longTranscriptWindow, shortFinalTurnWindow, overflowingRailWindow, sidebarLongSessionsWindow, disclosureOutputWindow, staleSessionsWindow, gitReviewWindow, gitReviewLargeWindow, artifactPaneWindow, localeSwitchWindow, planRemindersWindow. 69 tests / 69 launches -> 34 tests / 34 launches. Local (4 workers): 57.9s -> 47.3/43.2/40.8s across three green rounds. CI runs a single worker, where launch count dominates wall time.

jackwener · 2026-08-08T07:29:57Z

Review decision

Problem is real: function-scoped fixtures re-launch Electron for identical seeds; launch count dominates CI wall time.

Approach is correct: merge same-fixture assertions into ordered phases (count pins first, destructive last, turn-settle between sends), keep must-isolate cases (storage-root-conflict, permission-mode dual scenarios, staged-Skills draft restore).

Landing via maintainer branch because the fork head conflicts with main after #2478 presentation cleanup. Rebase drops reintroduced presentation journeys (MCP layout matrix, providers header geometry, settings back-icon rail) while keeping product journeys and dead-fixture removal.

See the rebased PR for the final tree.

jackwener · 2026-08-08T07:30:08Z

Landing via rebased #2489 (presentation pins already removed on main via #2478 are not reintroduced). Thanks @UncertaintyDeterminesYou4ndMe!

UncertaintyDeterminesYou4ndMe · 2026-08-08T07:21:11Z

Part of #2390 (Electron E2E half; the Storybook matrix reduction is #2481 per the issue's separate-PR requirement).

Problem

Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — and most of them re-launched and re-seeded exactly the state their file-mates had just built. quote-companion.spec.ts alone launched 12 windows over one identical seed; new-messages-indicator.spec.ts re-ran its six-message transcript seed from scratch for its second test.

Change

Tests that share one fixture and compatible assertions now run as consecutive phases of one journey. 69 tests / 69 launches → 34 tests / 34 launches. Every merge is annotated in-file with the ordering constraint that made it safe:

Count pins run first (e.g. 总消息数 === 1 phases, the blocked-Skill phase that pins zero turns and zero sessions).
Destructive phases run last (rename in settings-projects, the don't-ask-again checkbox that suppresses close confirmations for the rest of a window).
Turn-settle waits guard every phase boundary that sends — an Enter during a streaming turn silently becomes steering, which two merges surfaced immediately.

file	tests	file	tests
quote-companion	12 → 4	composer-skill-invocation	7 → 3
settings	6 → 2	mcp	5 → 2
composer-mention-token	4 → 1	send-message	4 → 2
providers	3 → 1	settings-projects	3 → 1
7 two-test files	2 → 1 each

Also removes 11 window fixtures no spec references (dead seeding code left behind by earlier test deletions, fixtures.ts 592 → 427 lines).

Deliberately not merged, and why

storage-root-conflict: the cold-start race is the subject; sharing a window would erase the precondition.
permission-mode-surface: its two tests boot two different fixture scenarios.
The staged-Skills draft-restoration test keeps its own window — observed, not guessed: merged into a shared window, its final send serialized raw /skill: text instead of chips, because concurrent session activity triggers exactly the external-value-change editor rebuild the test guards. That failure mode is worth a launch.
quote-companion's batch-close: needs the confirmation dialog the numbered-tab journey suppresses.
Cross-file candidates (the invocableSkillsWindow family spans 5 files; sandboxBoundaryWindow spans 2) are left as a maintainer decision — merging across files moves regression ownership between files, which is an organizational call, not a mechanical one.

Distinct contracts preserved

All IPC, focus, scrolling, remount, persistence, lifecycle, and animation assertions are unchanged — phases are the original test bodies with their assertions intact (three assertions were strengthened in passing: hostile-Mermaid now pins settled-turn count 2, the markdown --markdown scan asserts exit 0 explicitly, ask-user-question now answers after the reload, proving the rehydrated prompt is the live parked turn).

Timing

Same machine, warm build; CI runs workers=1 where launch count dominates:

config	before	after
workers=4 (local)	69 passed, 57.9s	34 passed, 47.3 / 43.2 / 40.8s (3 rounds)
workers=1 (CI-like)	69 passed, 192.4s	34 passed, 147.2s (−23%)

Electron launches: 69 → 34. Typecheck and biome are clean.

cc @Astro-Han

Part of apache#2390 (Electron E2E half; the Storybook matrix is a separate PR per the issue). Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — with most tests re-launching and re-seeding the exact state their file-mates had just built. Tests that share one fixture and compatible assertions now run as consecutive phases of one journey; every merge is annotated with the ordering constraint that made it safe (count pins run first, destructive phases run last, don't-ask-again phases end their window). Merged: ask-user-question 2->1, attachment 2->1, bot-onboarding 2->1, composer-mention-token 4->1, composer-skill-invocation 7->3, keyboard-help 2->1, mcp 5->2, new-messages-indicator 2->1 (also stops re-seeding the six-message transcript), providers 3->1, quote-companion 12->4, send-message 4->2, session-workbar 2->1, settings-projects 3->1, settings 6->2, skill-delete-scope 2->1. Kept separate deliberately: - storage-root-conflict (cold-start race is the subject), - permission-mode-surface (two different fixture scenarios), - the staged-Skills draft-restoration test (its contract is editor rebuilds on external value changes; concurrent session activity in a shared window perturbs exactly that — observed, not guessed), - quote-companion's batch-close (needs the confirmation dialog the numbered-tab journey suppresses via don't-ask-again). Also removes 11 window fixtures that no spec references (dead seeding code left behind by earlier test deletions): longTranscriptWindow, shortFinalTurnWindow, overflowingRailWindow, sidebarLongSessionsWindow, disclosureOutputWindow, staleSessionsWindow, gitReviewWindow, gitReviewLargeWindow, artifactPaneWindow, localeSwitchWindow, planRemindersWindow. 69 tests / 69 launches -> 34 tests / 34 launches. Local (4 workers): 57.9s -> 47.3/43.2/40.8s across three green rounds. CI runs a single worker, where launch count dominates wall time.

jackwener · 2026-08-08T07:29:57Z

Review decision

Problem is real: function-scoped fixtures re-launch Electron for identical seeds; launch count dominates CI wall time.

Approach is correct: merge same-fixture assertions into ordered phases (count pins first, destructive last, turn-settle between sends), keep must-isolate cases (storage-root-conflict, permission-mode dual scenarios, staged-Skills draft restore).

Landing via maintainer branch because the fork head conflicts with main after #2478 presentation cleanup. Rebase drops reintroduced presentation journeys (MCP layout matrix, providers header geometry, settings back-icon rail) while keeping product journeys and dead-fixture removal.

See the rebased PR for the final tree.

jackwener · 2026-08-08T07:30:08Z

Landing via rebased #2489 (presentation pins already removed on main via #2478 are not reintroduced). Thanks @UncertaintyDeterminesYou4ndMe!

UncertaintyDeterminesYou4ndMe · 2026-08-08T07:21:11Z

Part of #2390 (Electron E2E half; the Storybook matrix reduction is #2481 per the issue's separate-PR requirement).

Problem

Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — and most of them re-launched and re-seeded exactly the state their file-mates had just built. quote-companion.spec.ts alone launched 12 windows over one identical seed; new-messages-indicator.spec.ts re-ran its six-message transcript seed from scratch for its second test.

Change

Tests that share one fixture and compatible assertions now run as consecutive phases of one journey. 69 tests / 69 launches → 34 tests / 34 launches. Every merge is annotated in-file with the ordering constraint that made it safe:

Count pins run first (e.g. 总消息数 === 1 phases, the blocked-Skill phase that pins zero turns and zero sessions).
Destructive phases run last (rename in settings-projects, the don't-ask-again checkbox that suppresses close confirmations for the rest of a window).
Turn-settle waits guard every phase boundary that sends — an Enter during a streaming turn silently becomes steering, which two merges surfaced immediately.

file	tests	file	tests
quote-companion	12 → 4	composer-skill-invocation	7 → 3
settings	6 → 2	mcp	5 → 2
composer-mention-token	4 → 1	send-message	4 → 2
providers	3 → 1	settings-projects	3 → 1
7 two-test files	2 → 1 each

Also removes 11 window fixtures no spec references (dead seeding code left behind by earlier test deletions, fixtures.ts 592 → 427 lines).

Deliberately not merged, and why

storage-root-conflict: the cold-start race is the subject; sharing a window would erase the precondition.
permission-mode-surface: its two tests boot two different fixture scenarios.
The staged-Skills draft-restoration test keeps its own window — observed, not guessed: merged into a shared window, its final send serialized raw /skill: text instead of chips, because concurrent session activity triggers exactly the external-value-change editor rebuild the test guards. That failure mode is worth a launch.
quote-companion's batch-close: needs the confirmation dialog the numbered-tab journey suppresses.
Cross-file candidates (the invocableSkillsWindow family spans 5 files; sandboxBoundaryWindow spans 2) are left as a maintainer decision — merging across files moves regression ownership between files, which is an organizational call, not a mechanical one.

Distinct contracts preserved

All IPC, focus, scrolling, remount, persistence, lifecycle, and animation assertions are unchanged — phases are the original test bodies with their assertions intact (three assertions were strengthened in passing: hostile-Mermaid now pins settled-turn count 2, the markdown --markdown scan asserts exit 0 explicitly, ask-user-question now answers after the reload, proving the rehydrated prompt is the live parked turn).

Timing

Same machine, warm build; CI runs workers=1 where launch count dominates:

config	before	after
workers=4 (local)	69 passed, 57.9s	34 passed, 47.3 / 43.2 / 40.8s (3 rounds)
workers=1 (CI-like)	69 passed, 192.4s	34 passed, 147.2s (−23%)

Electron launches: 69 → 34. Typecheck and biome are clean.

cc @Astro-Han

Part of apache#2390 (Electron E2E half; the Storybook matrix is a separate PR per the issue). Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — with most tests re-launching and re-seeding the exact state their file-mates had just built. Tests that share one fixture and compatible assertions now run as consecutive phases of one journey; every merge is annotated with the ordering constraint that made it safe (count pins run first, destructive phases run last, don't-ask-again phases end their window). Merged: ask-user-question 2->1, attachment 2->1, bot-onboarding 2->1, composer-mention-token 4->1, composer-skill-invocation 7->3, keyboard-help 2->1, mcp 5->2, new-messages-indicator 2->1 (also stops re-seeding the six-message transcript), providers 3->1, quote-companion 12->4, send-message 4->2, session-workbar 2->1, settings-projects 3->1, settings 6->2, skill-delete-scope 2->1. Kept separate deliberately: - storage-root-conflict (cold-start race is the subject), - permission-mode-surface (two different fixture scenarios), - the staged-Skills draft-restoration test (its contract is editor rebuilds on external value changes; concurrent session activity in a shared window perturbs exactly that — observed, not guessed), - quote-companion's batch-close (needs the confirmation dialog the numbered-tab journey suppresses via don't-ask-again). Also removes 11 window fixtures that no spec references (dead seeding code left behind by earlier test deletions): longTranscriptWindow, shortFinalTurnWindow, overflowingRailWindow, sidebarLongSessionsWindow, disclosureOutputWindow, staleSessionsWindow, gitReviewWindow, gitReviewLargeWindow, artifactPaneWindow, localeSwitchWindow, planRemindersWindow. 69 tests / 69 launches -> 34 tests / 34 launches. Local (4 workers): 57.9s -> 47.3/43.2/40.8s across three green rounds. CI runs a single worker, where launch count dominates wall time.

jackwener · 2026-08-08T07:29:57Z

Review decision

Problem is real: function-scoped fixtures re-launch Electron for identical seeds; launch count dominates CI wall time.

Approach is correct: merge same-fixture assertions into ordered phases (count pins first, destructive last, turn-settle between sends), keep must-isolate cases (storage-root-conflict, permission-mode dual scenarios, staged-Skills draft restore).

Landing via maintainer branch because the fork head conflicts with main after #2478 presentation cleanup. Rebase drops reintroduced presentation journeys (MCP layout matrix, providers header geometry, settings back-icon rail) while keeping product journeys and dead-fixture removal.

See the rebased PR for the final tree.

jackwener · 2026-08-08T07:30:08Z

Landing via rebased #2489 (presentation pins already removed on main via #2478 are not reintroduced). Thanks @UncertaintyDeterminesYou4ndMe!

UncertaintyDeterminesYou4ndMe · 2026-08-08T07:21:11Z

Part of #2390 (Electron E2E half; the Storybook matrix reduction is #2481 per the issue's separate-PR requirement).

Problem

Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — and most of them re-launched and re-seeded exactly the state their file-mates had just built. quote-companion.spec.ts alone launched 12 windows over one identical seed; new-messages-indicator.spec.ts re-ran its six-message transcript seed from scratch for its second test.

Change

Tests that share one fixture and compatible assertions now run as consecutive phases of one journey. 69 tests / 69 launches → 34 tests / 34 launches. Every merge is annotated in-file with the ordering constraint that made it safe:

Count pins run first (e.g. 总消息数 === 1 phases, the blocked-Skill phase that pins zero turns and zero sessions).
Destructive phases run last (rename in settings-projects, the don't-ask-again checkbox that suppresses close confirmations for the rest of a window).
Turn-settle waits guard every phase boundary that sends — an Enter during a streaming turn silently becomes steering, which two merges surfaced immediately.

file	tests	file	tests
quote-companion	12 → 4	composer-skill-invocation	7 → 3
settings	6 → 2	mcp	5 → 2
composer-mention-token	4 → 1	send-message	4 → 2
providers	3 → 1	settings-projects	3 → 1
7 two-test files	2 → 1 each

Also removes 11 window fixtures no spec references (dead seeding code left behind by earlier test deletions, fixtures.ts 592 → 427 lines).

Deliberately not merged, and why

storage-root-conflict: the cold-start race is the subject; sharing a window would erase the precondition.
permission-mode-surface: its two tests boot two different fixture scenarios.
The staged-Skills draft-restoration test keeps its own window — observed, not guessed: merged into a shared window, its final send serialized raw /skill: text instead of chips, because concurrent session activity triggers exactly the external-value-change editor rebuild the test guards. That failure mode is worth a launch.
quote-companion's batch-close: needs the confirmation dialog the numbered-tab journey suppresses.
Cross-file candidates (the invocableSkillsWindow family spans 5 files; sandboxBoundaryWindow spans 2) are left as a maintainer decision — merging across files moves regression ownership between files, which is an organizational call, not a mechanical one.

Distinct contracts preserved

All IPC, focus, scrolling, remount, persistence, lifecycle, and animation assertions are unchanged — phases are the original test bodies with their assertions intact (three assertions were strengthened in passing: hostile-Mermaid now pins settled-turn count 2, the markdown --markdown scan asserts exit 0 explicitly, ask-user-question now answers after the reload, proving the rehydrated prompt is the live parked turn).

Timing

Same machine, warm build; CI runs workers=1 where launch count dominates:

config	before	after
workers=4 (local)	69 passed, 57.9s	34 passed, 47.3 / 43.2 / 40.8s (3 rounds)
workers=1 (CI-like)	69 passed, 192.4s	34 passed, 147.2s (−23%)

Electron launches: 69 → 34. Typecheck and biome are clean.

cc @Astro-Han

Part of apache#2390 (Electron E2E half; the Storybook matrix is a separate PR per the issue). Every e2e window fixture is function-scoped, so 69 tests paid 69 Electron launches — with most tests re-launching and re-seeding the exact state their file-mates had just built. Tests that share one fixture and compatible assertions now run as consecutive phases of one journey; every merge is annotated with the ordering constraint that made it safe (count pins run first, destructive phases run last, don't-ask-again phases end their window). Merged: ask-user-question 2->1, attachment 2->1, bot-onboarding 2->1, composer-mention-token 4->1, composer-skill-invocation 7->3, keyboard-help 2->1, mcp 5->2, new-messages-indicator 2->1 (also stops re-seeding the six-message transcript), providers 3->1, quote-companion 12->4, send-message 4->2, session-workbar 2->1, settings-projects 3->1, settings 6->2, skill-delete-scope 2->1. Kept separate deliberately: - storage-root-conflict (cold-start race is the subject), - permission-mode-surface (two different fixture scenarios), - the staged-Skills draft-restoration test (its contract is editor rebuilds on external value changes; concurrent session activity in a shared window perturbs exactly that — observed, not guessed), - quote-companion's batch-close (needs the confirmation dialog the numbered-tab journey suppresses via don't-ask-again). Also removes 11 window fixtures that no spec references (dead seeding code left behind by earlier test deletions): longTranscriptWindow, shortFinalTurnWindow, overflowingRailWindow, sidebarLongSessionsWindow, disclosureOutputWindow, staleSessionsWindow, gitReviewWindow, gitReviewLargeWindow, artifactPaneWindow, localeSwitchWindow, planRemindersWindow. 69 tests / 69 launches -> 34 tests / 34 launches. Local (4 workers): 57.9s -> 47.3/43.2/40.8s across three green rounds. CI runs a single worker, where launch count dominates wall time.

jackwener · 2026-08-08T07:29:57Z

Review decision

Problem is real: function-scoped fixtures re-launch Electron for identical seeds; launch count dominates CI wall time.

Approach is correct: merge same-fixture assertions into ordered phases (count pins first, destructive last, turn-settle between sends), keep must-isolate cases (storage-root-conflict, permission-mode dual scenarios, staged-Skills draft restore).

Landing via maintainer branch because the fork head conflicts with main after #2478 presentation cleanup. Rebase drops reintroduced presentation journeys (MCP layout matrix, providers header geometry, settings back-icon rail) while keeping product journeys and dead-fixture removal.

See the rebased PR for the final tree.

jackwener · 2026-08-08T07:30:08Z

Landing via rebased #2489 (presentation pins already removed on main via #2478 are not reintroduced). Thanks @UncertaintyDeterminesYou4ndMe!

UncertaintyDeterminesYou4ndMe mentioned this pull request Aug 8, 2026
perf(test): remove redundant desktop verification setup #2390
Closed

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

jackwener closed this Aug 8, 2026

UncertaintyDeterminesYou4ndMe mentioned this pull request Aug 8, 2026
perf(test): remove redundant desktop verification setup #2390
Closed

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

jackwener closed this Aug 8, 2026

UncertaintyDeterminesYou4ndMe mentioned this pull request Aug 8, 2026
perf(test): remove redundant desktop verification setup #2390
Closed

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

jackwener closed this Aug 8, 2026

UncertaintyDeterminesYou4ndMe mentioned this pull request Aug 8, 2026
perf(test): remove redundant desktop verification setup #2390
Closed

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

jackwener closed this Aug 8, 2026

UncertaintyDeterminesYou4ndMe mentioned this pull request Aug 8, 2026
perf(test): remove redundant desktop verification setup #2390
Closed

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

jackwener closed this Aug 8, 2026

UncertaintyDeterminesYou4ndMe mentioned this pull request Aug 8, 2026
perf(test): remove redundant desktop verification setup #2390
Closed

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

jackwener closed this Aug 8, 2026

UncertaintyDeterminesYou4ndMe mentioned this pull request Aug 8, 2026
perf(test): remove redundant desktop verification setup #2390
Closed

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

jackwener closed this Aug 8, 2026

UncertaintyDeterminesYou4ndMe mentioned this pull request Aug 8, 2026
perf(test): remove redundant desktop verification setup #2390
Closed

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

jackwener closed this Aug 8, 2026

Uh oh!

Conversation

UncertaintyDeterminesYou4ndMe commented Aug 8, 2026

Problem

Change

Deliberately not merged, and why

Distinct contracts preserved

Timing

Uh oh!

jackwener commented Aug 8, 2026

Review decision

Uh oh!

jackwener commented Aug 8, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

UncertaintyDeterminesYou4ndMe commented Aug 8, 2026

Problem

Change

Deliberately not merged, and why

Distinct contracts preserved

Timing

Uh oh!

jackwener commented Aug 8, 2026

Review decision

Uh oh!

jackwener commented Aug 8, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

UncertaintyDeterminesYou4ndMe commented Aug 8, 2026

Problem

Change

Deliberately not merged, and why

Distinct contracts preserved

Timing

Uh oh!

jackwener commented Aug 8, 2026

Review decision

Uh oh!

jackwener commented Aug 8, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

UncertaintyDeterminesYou4ndMe commented Aug 8, 2026

Problem

Change

Deliberately not merged, and why

Distinct contracts preserved

Timing

Uh oh!

jackwener commented Aug 8, 2026

Review decision

Uh oh!

jackwener commented Aug 8, 2026

Uh oh!