desktop: native macOS toolbar, consistent page design, and snappier navigation

[h3nock]

What

Reworks the desktop app chrome and page design language toward native macOS, and cuts navigation waste. Three themes:

1. Native macOS unified toolbar. Replaces the in-content header chrome with a real NSToolbar (.windowToolbarStyle(.unified)): traffic lights, a Home affordance + page title on the left, and Capture / Listening / Settings on the right — all in the system title bar. Capture and Listening are labeled with a readable on/off state (green-tinted when running, amber when blocked, dimmed when off), not a tiny badge. Toolbar glyphs stay neutral (no purple). The window content now runs edge-to-edge under the toolbar instead of sitting in a nested rounded card.

2. Consistent page design language. A shared header-control kit (OmiHeaderControls) unifies the search field, filter chips, and icon buttons across Conversations, Tasks, Memories, and Rewind — one capsule search, one chip style, one circular icon button, one active-state treatment. Replaces the previous mix of radius-8 squares, 42pt white squares, and assorted chip heights. Task rows are cleaned up (detail button moved into the hover cluster instead of a permanent badge line, checkbox aligned to the first text line). The chat input collapses from a field-inside-a-panel into a single surface.

3. Snappier navigation. Pages were recreated on every tab switch, so page-held state and TTLs died per visit (Home refetched all status data every visit; the brain map re-ran an 800-tick force layout and rebuilt the SceneKit scene every Memories visit). Hoisted into session-persistent owners in ViewModelContainer:

• HomeStatusStore — counts/connector/MCP statuses behind a 60s activation throttle + single-flight
• ConversationsSearchModel — query + debounce + results survive navigation
• shared MemoryGraphViewModel — scene persists; skips re-simulation when the graph content-signature is unchanged; on-disk layout cache (FNV-1a keyed) restores a settled layout in ~1ms vs a multi-thousand-ms cold sim

Verification

• xcrun swift build -c debug --package-path Desktop
• xcrun swift test -c debug --package-path Desktop --filter DashboardCaptureStateTests
• xcrun swift test -c debug --package-path Desktop --filter ImportConnectorStatusStoreTests
• xcrun swift test -c debug --package-path Desktop --filter ScreenRecordingPermissionPolicyTests
• python3 scripts/check-e2e-flow-coverage.py --strict --base upstream/main
• python3 scripts/check-userdefaults-key-ratchet.py --print
• git diff --check
• Rebased onto current upstream/main; conflict resolved by keeping the shared toolbar controller and porting upstream's screen-recording permission helper into it.

[cubic-dev-ai]

4 issues found across 23 files

Confidence score: 3/5

• desktop/macos/Desktop/Sources/Stores/HomeStatusStore.swift has a race where resetSessionState() can run while an in-flight refresh() resumes and writes stale per-account data, so account switches/sign-out can show the wrong status for the new session — add cancellation/epoch guards so refresh() results are ignored after reset before merging.
• desktop/macos/Desktop/Sources/Stores/HomeStatusStore.swift also leaves importConnectorStatusStore intact on reset, which can leak connector badge state across users and show incorrect connected indicators after sign-out/switch — clear or reinitialize that nested store in resetSessionState() to keep session boundaries clean.
• desktop/macos/Desktop/Sources/MainWindow/Pages/MemoryGraph/MemoryGraphPage.swift consumes hasRunEmptyBootstrap before rebuildGraph() succeeds, so one transient failure can disable empty-graph bootstrap for the rest of the session and leave the page stuck without expected recovery — only flip the latch after a successful rebuild (or retry/fallback on failure).
• desktop/macos/Desktop/Sources/MainWindow/DesktopHomeView.swift applies unified toolbar styling even when useLegacyHomeDesign hides mainToolbar, which can yield an empty toolbar in legacy mode and a visibly broken header — gate the unified style with the same flag or provide a legacy toolbar fallback before merging.

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

2 issues found across 17 files (changes from recent commits).

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

2 issues found across 4 files (changes from recent commits).

<sub>Tip: Review your code locally with the cubic CLI to iterate faster.

Re-trigger cubic</sub>

[Git-on-my-level]

We like most of the work in this PR: the faster page navigation, persistent Home/search state, memory graph improvements, page consistency, and cleanup all look directionally good. The native macOS toolbar/nav treatment does not feel on-brand for Omi, though, so we plan to revert that piece while preserving the rest of the improvements.

[Git-on-my-level]

Pushed a follow-up commit to keep the PR’s non-toolbar improvements while moving the nav/status chrome back into the app surface.

Remaining feedback:

• The nav/header still is not fully restored or polished enough yet. The native macOS toolbar direction was not on-brand, and the in-app restoration still needs UI/UX tuning before merge: spacing, title/back affordance, status-control placement, and page-to-page consistency all need another pass.
• The footgun level is better, but still not low enough. The PR moves more state to longer-lived owners, which is directionally good, but the new stores/controllers still lean on singletons, UserDefaults, NotificationCenter, and silent try? fallbacks. That leaves too many paths where failures can become stale UI instead of explicit, testable states.
• Testing is still thin for the amount of behavior being moved. The current coverage is mostly source-inspection tests plus a few UserDefaults/session-scope tests. We should add more behavioral coverage around search cancellation/order, Home refresh failure/freshness states, memory graph runtime behavior, and manual or automation-backed visual QA for the page chrome.

So overall: we like most of the work and should preserve the performance/state improvements, but this is not yet at a low-footgun, easy-to-maintain bar until the remaining chrome polish and behavioral tests are tightened.

[Git-on-my-level]

Thanks for the follow-up here — the non-toolbar performance/state work is directionally useful, but I found one blocker before this is merge-ready:

• ViewModelContainer.resetStartupState() resets the Home/search/task/memory/app stores on account changes, but it does not reset or replace the session-persistent memoryGraphViewModel. Since this PR intentionally makes the brain-map scene/layout long-lived, signing out or switching accounts can leave the previous user's memory graph nodes/edges and SceneKit scene visible to the next account until a later refresh replaces it. That is a cross-account stale-data/privacy leak, not just a freshness issue. Please add an explicit reset path for MemoryGraphViewModel (clear scene nodes/edges, simulation, loaded signature/cache state, loading/bootstrap flags, selected node, etc.) and cover it with a behavioral account-switch/session reset test.

Keeping this as request-changes because it affects user memory data isolation. I agree with the existing maintainer direction that the remaining chrome polish and behavioral coverage should also be tightened before merge.