You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Spike: native Mac and Windows keyboard predictions #797
Prototype native word prediction in the Mac and Windows scanning keyboard, using native results by default and the existing database as fallback. Depends on #795 / #796; keep delivery in a separate draft PR until spike findings are reviewed.
Use NSSpellChecker completion with surrounding text and prefix ranges on Mac, and TextPredictionGenerator on Windows. Prefer UK English; retry US English on Windows when UK yields no usable results, and generic English on Mac if UK is unavailable. Preserve ranking, bound/filter candidates, retain suffix-only insertion and context/cancellation safeguards. Native timeout/failure/empty results fall back to the database within the existing two-second worker deadline. No saved schema, BLE or cloud-service changes.
Validate fake provider ranking/filtering/fallback/deadlines/cancellation/locale retry/exactly-once insertion; stable scanning and repeated predictions; synthetic comparisons and native switch typing in both VMs, including offline, protected fields, selections and focus changes. Restore VM state and leave the local signed Mac app untouched. Run frontend/Rust checks, independent latest-head review and CI. No merge or release is authorized.
Spike findings
Draft #798 is stacked on #796. Measured results and reproduction include 120 successful synthetic native requests across both VMs online/offline, with no fallback in that sample. Native-first generation, bounded helper cleanup and fake-provider coverage are implemented.
Production qualification remains blocked: Mac switch-driven letters, Space and Close work, but TextEdit's suggestion row remained empty; the shared context/worker/provider path has not yet been isolated. Windows interactive computer control was unreliable even with Switchify stopped. Live suggestion acceptance, row timing and the full protected/focus/selection matrix are not passed. Keep both providers experimental and the PR draft. VM settings/networks were restored; the signed local Mac app was untouched.
Activity