Widget publishes native input state per tab

[malmstein]

Task/Issue URL: https://app.asana.com/1/137249556945/project/1157893581871903/task/1214790107756091?focus=true

Stacked on top of #8549.

Description

PR 2 of the native input state refactor — see the parent task for the full plan.

The native input widget is now the single publisher of NativeInputState and reads its own state from the per-tab store (NativeInputStateProvider) instead of keeping it as a widget-local field. Plugins still won't see this change until PR 3.

Concretely:

• NativeInputModeWidgetViewModel injects NativeInputStatePublisher. A new init { … } block collects combine(state, activeTabId) and pushes each value through publisher.publish(tabId, state.copy(tabId = tabId)) whenever a tabId is set.
• NativeInputModeWidget injects NativeInputStateProvider. observeNativeInputState() now subscribes to provider.stateForTab(tabId) instead of viewModel.state. getInputState() reads provider.stateForTab(activeTabId).value.
• configure() and configureContextual() take a tabId: String parameter and start the observation once the tabId is known.
• tabId is threaded:
  • BrowserTabFragment.tabId → NativeInputManager.showNativeInput(tabId, …) → attachWidget(widgetView, isBottom, tabId) → widget.configure(tabId, …).
  • DuckChatContextualFragment reads its KEY_DUCK_AI_CONTEXTUAL_TAB_ID argument once and passes it to ContextualNativeInputManager.init(tabId, …) → widget.configureContextual(tabId).
• viewModel.state stays as a SharedFlow<NativeInputState> so existing ViewModel-level tests can keep asserting on its values; the widget no longer consumes it.

What this PR intentionally does not do:

• NativeInputState.tabId stays nullable. Tightening to non-null will follow once the publish-on-configure invariant is verified in practice (and we can drop the zero("__init__")-style placeholder).
• NativeInputHost.getInputState() is still exposed — its implementation now reads from the provider, but the surface remains until plugins are migrated in PR 3 and it's removed in PR 4.
• Plugins still receive the widget as their NativeInputHost; nothing about their behaviour changes here.

Steps to test this PR

• CI green: :duckchat-impl:testDebugUnitTest, :app:testPlayDebugUnitTest, spotlessCheck.
• Existing NativeInputModeWidgetViewModelTest cases still pass after the configure/configureContextual signature changes and the added publisher mock.
• Smoke-test on device:
  • Open Duck.ai input from the NTP, type, switch to the chat tab, submit — behaviour matches develop.
  • Open the omnibar on a browser tab, switch between Search and Duck.ai modes — toggle/position/context behave as before.
  • Open Duck.ai contextual chat from a browser tab — keyboard, attachments, and submit all still work.
  • Background a tab and reopen it — input state for that tab is restored correctly.

UI changes

None.

---

Note

Medium Risk
Changes how the native input widget publishes/consumes state by keying everything on tabId, which could cause cross-tab state leaks or missing updates if tab IDs aren’t consistently threaded through all entry points.

Overview
The native input widget now publishes NativeInputState per tab and reads state from NativeInputStateProvider instead of relying on widget-local/ViewModel flow state.

This threads tabId through BrowserTabFragment → NativeInputManager.showNativeInput/attachWidget → NativeInputModeWidget.configure, and through contextual chat via DuckChatContextualFragment/ContextualNativeInputManager.init → configureContextual. The ViewModel now injects NativeInputStatePublisher to push updates for the active tab, and tests are updated for the new tabId-required configure APIs.

^{Reviewed by Cursor Bugbot for commit 9dc03e8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Stale state causes applyOmnibarShape to permanently skip

Medium Severity

The old configure() synchronously initialized nativeInputState from viewModel.state.replayCache.lastOrNull(), which reflected the real computed inputMode. The new code replaces that with observeNativeInputState(), which reads from NativeInputStateProvider. The provider's StateFlow contains NativeInputState.zero (with inputMode = SEARCH_AND_DUCK_AI, hence toggleVisible = true) until the ViewModel's async coroutine chain publishes the real state. applyOmnibarShape is called immediately after and checks nativeInputState.toggleVisible — the zero state's true causes an early return, skipping shape application. Since applyState() never calls applyOmnibarShape, the shape is permanently wrong for users in SEARCH_ONLY mode.

Additional Locations (1)

• duckchat/duckchat-impl/src/main/java/com/duckduckgo/duckchat/impl/ui/nativeinput/views/NativeInputModeWidget.kt#L766-L780

 

^{Reviewed by Cursor Bugbot for commit 20634c4. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

There are 2 total unresolved issues (including 1 from previous review).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit e8d711f. Configure here.}

[GerardPaligot]

Work as expected. I tried to reproduce what is reported by Bugbot but I didn't find a way to reproduce. Good for me.

[YoussefKeyrouz]

Not sure if this could happen, but if the argument is missing, then we will be storing the data under the “” key. So any contextual fragment that might hit this path will share the same stored map entry. Should be have any null enforcement here?