ADFA-5132 | Prevent silent fallback to remote LLM providers

[jatezzz]

Description

This PR prevents the AI chat from silently routing user prompts to a remote provider when the selected local backend is unconfigured.

• What: Requests now strictly remain on the selected backend; if it is not ready, the system halts the request and notifies the user.

• How: Modified LlmInferenceServiceImpl.kt and ChatViewModel.kt to resolve the effective backend strictly based on the stored selection, completely bypassing the fallback to other available backends. Introduced a SelectedBackend sealed interface (Installed, Missing, None) to accurately identify and report the backend's exact status.

• Why: Users choosing local models expect their prompts and source code to remain strictly on-device. Silent fallbacks compromise privacy by sending data to third parties without consent, while also unexpectedly consuming API quotas.

Details

• Updated strings.xml with clear error messages, such as explicitly stating that a specific backend is not ready and confirming that nothing was sent.

• The system now correctly provides distinct fixes for the user depending on whether they need to configure their chosen backend, reinstall its missing plugin, or install any backend at all.

Before

document_4907021400063084965.mp4

After

document_4907021400063084971.mp4

Ticket

ADFA-5132

Also fixes:
ADFA-5133
ADFA-5134

Observation

To prevent race conditions where a superseded availability check could overwrite a newer verdict, the check logic in ChatViewModel.kt was refactored. It now utilizes a backendCheckSequence number confined to the main thread to ensure only the most recent status is published.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}