ClipLingo explores a low-interruption translation interaction for Windows: select English text in another application, press a global shortcut, and read an Indonesian translation near the cursor.
The Rust/Tauri host owns orchestration, while Win32 APIs provide controlled text capture and a lightweight popup. An optional local CTranslate2 build runs the EN-to-ID model without a cloud translation request.
Text and inference boundaries
Capture is permitted only for a current, authorized foreground source. A monotonically increasing request generation prevents older asynchronous translation results from replacing newer ones. Pending requests are bounded; neither the selected source text nor the translated output is intentionally retained as a translation history.
The model is loaded from a validated local pack. Multi-sentence requests are broken into at most eight segments with token limits, and incomplete output is rejected rather than silently presented as a full translation.
Native popup
The presentation remains target-only and uses non-activating Win32 flags. F2.9 adds a measured viewport and mouse-wheel scroll state for text longer than the popup height while retaining the complete bounded result. Deterministic tests check scroll limits, monitor coordinates including negative ones, and reset behavior.
Result and limits
The implementation has Windows CI and real headless model regression evidence. It is not yet a production-ready or publicly released translator. End-to-end desktop focus behavior, inactive-window wheel routing, screen readers, mixed-DPI text shaping, and practical latency still require tests on real Windows systems.