Repository navigation
Native key input: per-platform status (v1.7 attempt-first /key) #3
Description
Activity
macOS: RESOLVED in b3c24a0 — focus-gated CGEvents, verified in-page
The activation story that unblocked it:
show_window()is the whole recipe — it materializes the ghost's window, converts it to a titled window (the borderless ghost mask can never become key), centers it, activates the app, and makes it key. Only whenisKeyWindowconfirms does anything get posted — posting into the HID tap without focus would type into someone else's app.Layout independence: single characters carry
CGEventKeyboardSetUnicodeString— keycode 0x00 is 'a' on QWERTY but 'q' on AZERTY; the attached unicode string wins, so the agent's character is what lands ('Z' arrives uppercase with no shift needed).Honesty preserved: a document-level once-listener must observe
{trusted:true}before Ok is reported — otherwise Err and the synthetic fallback runs with the mode saying so.Live-tested on AZERTY hardware:
a,Enter,Z,Tab,ArrowLeft→mode: native,TRUSTED:<key>on all five, zero fallbacks.Still open here: Wayland (XTEST is X11-only; fine under XWayland, needs libei/uinput on pure Wayland) and keymap coverage growth (the mac table covers letters/digits/punctuation/named keys; the X11 keysym table can grow the same way).
Closing with the scoreboard: macOS — RESOLVED and live-verified (v1.9.0, AZERTY hardware 5/5 TRUSTED, show_window focus gate + CGEventKeyboardSetUnicodeString, b3c24a0). Windows — verified end-to-end since v1.7 (SendInput). Remaining: Wayland (XTEST is X11-only; needs libei/uinput) and keymap table growth — demoted to backlog, this issue is the record; reopen when Wayland lands on the roadmap. The synthetic fallback carries every gap in the meantime, with the mode honestly reported.
v1.7 shipped
POST /keyas attempt-first, honest mode: real OS-level events when possible (isTrusted: true), automatic fallback to the in-page synthetic dispatch, and the response always reports whichmoderan. The README links here for where each platform stands.Status per platform
SendInput+SetForegroundWindow, targeting the WebView2 child HWND viaEnumChildWindows. Native events land in the page; this is the reference implementation.x11rbXTESTfake_inputwith keycode resolution from the live keymap. The ghost window is brought on-screen first (X11 refuses focus to unviewable windows), and deliberately nogtk::main_iteration()runs on the HTTP thread (GTK is main-thread-only; pumping from a worker deadlocked the loop — that was CI exit 52). Headless CI has no focus target → falls back cleanly.document.hasFocus() == 0. Until that's solved, macOS reports "native unavailable" and falls back to synthetic.Open work
session_show-assisted native typing on the visible window)x11_keysymtable; special keys/modifiers on all three paths