Lumi AI changelog

0.8.5

Fixed Improvement Center actions returning HTTP 404 because their hidden action field could shadow the form's endpoint URL in the browser.
Reused Lumi's shared clobber-safe form resolver for feedback actions and asynchronous Lumi AI settings/download forms.

Fixed no-change, dismiss, edit, and restore actions in the Improvement Center by using an explicit authenticated JSON response contract.
Replaced the generic update failure with actionable session, permission, missing-record, HTTP, and server-write messages.
Added safe server diagnostics and regression coverage for the affected feedback state transitions.

Replaced the multi-step approve, promote, and save workflow with a guided one-step “Apply improvement” action that activates reviewed corrections immediately.
Added direct dismiss and no-change decisions, automatically moving all finalized feedback out of the active queue.
Added a searchable feedback-history modal with outcome filters, full record inspection, restoration, editing, and timed permanent deletion.
Added live queue synchronization for new feedback and reviewer actions, including clean reconnect behavior without full-page refreshes.
Moved correction-bank, evaluation, export, and reviewer-access controls into a clearly labeled advanced area.

Added OKF-first, role-aware retrieval with compact fast-model answers, conversational subject resolution, deterministic verified links, and main-model fallback when local evidence is insufficient.
Added controller-directed response modes, retrieval depth, output sizing, lifecycle metrics, grouped work history, and richer diagnostics.
Hardened Web Search tool exposure and source grounding so external lookup is used only when relevant local knowledge is insufficient.
Improved runtime/model installation, hardware guidance, failure recovery, tool management, feedback/correction workflows, and settings usability.
Preserved local models, runtimes, settings, metrics, feedback, and tool data during updates and snapshots.