Skip to content

Instantly share code, notes, and snippets.

@mxschmitt
Last active April 1, 2026 23:46
Show Gist options
  • Select an option

  • Save mxschmitt/ad4cdded629b61f3af43ad12e796704e to your computer and use it in GitHub Desktop.

Select an option

Save mxschmitt/ad4cdded629b61f3af43ad12e796704e to your computer and use it in GitHub Desktop.
Replay Engine PoC - All-Hands April 4, 2026
marp true
theme default
paginate true

Nova Act Realtime Pod

First update

An agent that keeps up with you. Sees what you see. Adjusts when you adjust.

This week: We proved the data we record is correct -- and that we can act on it.


What we built

Record raw OS events (mouse, keyboard, scroll, touchpad gestures) into SQLite Replay them 1:1 at microsecond timing to drive a real browser Verify with a 23-level E2E test suite that compares DOM state

This is the actuation layer for the agent -- the foundation for it to take actions in the real world.

Cross-platform: macOS (Quartz) + Windows (SendInput)


macOS Demo

What you're seeing:

  • Real user session recorded, then replayed into the same browser
  • Clicks land, text appears, scroll positions match
  • The browser can't tell the difference between the user and the replay

Narrate live over the video


Windows Demo

What you're seeing:

  • 23 test levels: record, replay, compare -- pass/fail badges in real-time
  • 20 out of 23 tests pass

Narrate live over the video


What works, what doesn't

20/23 tests pass

Clicks, right-click, double/triple/middle-click, drag, text input, special keys, auto-repeat, copy/paste, shift+click, scroll direction, multi-line input, focus order, click accuracy (7/7 targets), form controls, drag-and-drop

2 hard problems identified

Problem Detail
Touchpad scroll Windows touchpad gestures go through an OS gesture recognizer that's non-deterministic -- replay produces different velocity each time. We investigated a kernel driver approach but it's too fragile to ship.
Mouse velocity 3.4% distance loss from OS acceleration + browser event coalescing at high speed

Key takeaway: We know exactly what works, what doesn't, and why. Touchpad gesture replay is a known unsolved problem industry-wide -- we're not shipping a fragile workaround.


Next steps

Touchpad replay -- parked

We investigated deeply (custom kernel driver, HID injection). Conclusion: Windows touchpad gesture replay is too fragile and non-deterministic to ship. No one in the industry has solved this reliably. We'll work around it at a higher level (e.g. scroll via SendInput instead of raw gestures).

Action optimization

Use the replay engine to take a recording and find the minimal action sequence that achieves the same result. The actuation layer + action space lets us treat any recording as a "gym pass" and discover the shortest path to the same outcome.

Nova Act Realtime vision

Record (what the user did) + Understand (what mattered) + Act (replay/optimize) The actuation layer we proved this week is the "Act" piece.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment