[broven]

Summary

Stacked on #133: review and merge that one first.

This PR adds simulated human input to executor click() / fill(). It never activates a tab and never focuses a window: the extension backend drives the user's own Chrome, so focus stealing interrupts them.

• repl/human_input.py, installed in _Worker._bind_live_surface. Locator, Page and Frame click/fill, plus Mouse wrappers:
  • pointer approach that samples many points on the way to a random point inside the target;
  • randomized press and key timing, and Shift chords;
  • fill() replaces existing text with a physical Ctrl/Cmd+A then Backspace;
  • targets outside the viewport are scrolled into view with native wheel input.
• Facade. Consecutive mouse moves are forwarded together and finished before any press, so Chrome delivers real coalesced pointer events. Generated clicks are snapped to the device-pixel grid when they are dispatched.
• Extension. It reports Chrome's real user agent through Browser.getVersion. Without this, Playwright assumes Linux on a Mac and presses Ctrl+A, which does not select all on macOS.
• No focus stealing. The battle branch's bring_to_front() calls and its Page.bringToFront / Target.activateTarget → chrome.tabs.update({active}) / chrome.windows.update({focused}) handling are not ported. The battle branch is the red-vs-blue detection exercise these changes come from.

Defaults, opt-outs and cost

• On by default on both backends. Pixel-grid snapping is extension-only.

• Opt out per call with human=False, or globally with BW_HUMAN_INPUT=0 in the daemon environment.

• Latency is significant. Measured on a headful background tab:

  Action	Time
  click()	~4 s
  9-character fill()	~3 s
  click that needs a wheel scroll	~5.4 s

  Native Playwright takes under 1 s for each. Details are in the "Cost" section of docs/human-input.md.

• An agent that explicitly calls page.bring_to_front() still gets Chrome's native behaviour. This is unchanged from main and documented.

Test plan

• mise run lint

• mise run test: skill 338; daemon 982 passed, 102 skipped; evals 13/13; pi.

• Real-Chrome E2E, 28 files: 63 passed. These are the new test_human_input.py, test_human_scroll.py, test_human_pointer_layout.py and test_human_input_no_focus.py, plus the Keep the extension backend quiet: isolated framework state, no Runtime/Debugger, no title marker #133 suites.

• test_human_input_no_focus.py runs click, fill and a wheel-scroll click on a background session tab, while another tab is active and another window has focus. It asserts:

  • every event on the page is trusted;
  • the active tab in every window and the focused window are unchanged;
  • no tabs.onActivated, tabs.onHighlighted or windows.onFocusChanged fired.

  As a sanity check, temporarily adding one bring_to_front() makes this test fail.

• The diff adds no bring_to_front, bringToFront, activateTarget, windows.update or tabs.update( call outside the docs.

🤖 Generated with Claude Code