fix(cua-driver): repair desktop discovery, typing and click delivery

[tanishqkancharla]

OpenSky could miss installed apps, misdirect desktop input, and submit a LibreOffice dialog when an agent clicked an editable field. This draft updates the fork’s OpenSky Driver identity and repairs observed app-discovery and input behavior.

The Linux adapter preserves virtual-table indices and screenshot geometry, resolves indexed clicks against the requested window, and uses pointer input where an accessibility activation is not equivalent to a click. Editable fields now follow that rule: clicking places the caret instead of invoking an activate action that can submit the dialog. Only proven pre-input refusal permits fallback; uncertain action failures are never replayed. Background typing retains the requested window and refuses a different active dialog before entering text.

The indexed-click baseline 4863e5bff14a98d349f1b16d803c9f65449dcbc6 passed compilation and focused Rust checks, then 12 real public-SDK cases covering indexed/scrolled cells, saved ranges, cross-window refusal, dialogs, screenshots and input. Saved outputs, owned app exits and keyboard restoration were independently checked.

The editable-click correction was isolated on that baseline as fbf0d0c632763f0daae0e326c1d2fe3a44a89f2a. Its build and 64 focused Rust tests passed. The same real LibreOffice test failed before the change because indexed clicking alone closed the dialog, then passed on the candidate: the dialog stayed open and the replacement title was saved for all 28 cells with choices and original values unchanged. Seven further SDK regressions passed: CALC-L01/L02/L05, FOCUS-L02/L03/L04 and DROPDOWN-L01. Independent readers checked all saved formulas, both documents in each focus test, and dropdown rules; all ten regression app groups exited and no container remained. Test source and validation record retain the scope and artifact identities.

The combined draft also includes bounded metadata-read concurrency and earlier focus/coordinate input improvements. Metadata concurrency was excluded from the isolated editable-click comparison. The frozen agent campaign still uses SDK 8c818c7 and driver 4863e5bff; no failed campaign result was replaced and no candidate performance gain is claimed here.

Native Wayland lacks this editable-field pointer fallback: the correction refuses activation before input but does not provide working indexed editable clicks there. Other limitations include conservative refusal when a same-process sibling owns focus, the existing global Tk fallback target, and the X11 check-to-input race. The combined draft has not completed the canonical desktop matrix, remaining macOS GUI checks, Windows lifecycle checks or full agent campaign. It remains draft, with no cross-platform parity or release readiness claim.

Other included work covers exact macOS installation/process association, tool roster and close-window bindings, real-driver fixtures, preserving Windows Computer History unless uninstall purge is explicit, and the upstream ScreenCaptureKit namespace correction. Contributor history is retained.