Repository navigation
Add focus guard (expectWindow) to keyboard input tools #12
Description
keyboard_type, keyboard_press, and keyboard_shortcut send input to whatever window currently has focus, with no way to assert what that window is.
Why this matters
During a playtest I had to hand-roll the safety check: focus_window returned false, so before typing I called get_active_window manually to confirm Notepad was actually foreground. Without that manual step the keystrokes would have gone into the user's editor — which had another agent actively working in it.
That check should be structural, not something each caller remembers. Frontier computer-use harnesses treat "focus changed unexpectedly -> stop typing" as a deterministic reflex, not a prompt-level instruction.
Fix
Add an optional expectWindow parameter (substring match on title, or a pid) to the keyboard tools. When supplied:
- Read the current foreground window.
- If it does not match, refuse to send input and return an error naming the window that actually had focus.
- If it matches, proceed.
Default behavior (no expectWindow) is unchanged, so this is backwards compatible. Optionally expose the same guard for mouse_click.
This is the first reflex in the server: a deterministic precondition check that does not require a model turn.
Activity