Skip to content

OpenMausBot 0.1.15

Choose a tag to compare

View all tags
@aivsomkar aivsomkar released this 13 Aug 19:00
v0.1.15
436271a

Computer use is about twice as fast

Every step a bot took on its cloud computer used to cost two model round trips — act, then a separate screenshot to see what happened — plus several sequential requests to the machine. Now the action, the settle and the screenshot all happen in one round trip, and the resulting screen comes back attached to the action itself.

Measured against a live box: one UI step went from ~11.7s to ~5.9s, and screenshots are ~43KB JPEGs instead of multi-hundred-KB PNGs.

  • Actions return the screen. click, type_text, press_key, scroll and open_url each return a fresh frame, so the agent never needs a follow-up screenshot.
  • computer_batch runs a whole predictable sequence — click a field, type, Tab, type, Return — in a single round trip with one screenshot at the end.
  • Sleeping computers wake themselves. Boxes archive when idle; actions now resume the machine and retry instead of failing.
  • Frames are verified whole before they reach the model, so a truncated capture can never arrive as a half-rendered image.
  • Click coordinates scale correctly on displays narrower than the screenshot space, failed actions are reported instead of silently swallowed, and batched steps get a beat between them so typing doesn't lose its first characters.
  • The live screen preview no longer competes with the bot for the machine, and bots whose engine can't drive a computer are no longer told they have one.

Box agent

  • The reply streams as it arrives instead of repeating the whole message on every update.
  • Turns settle correctly when the machine omits a prompt id, instead of hanging until the timeout.

Windows

  • Shell handling fixes and updated release documentation.

Install: download the .dmg (Apple silicon) or -setup.exe (Windows). Existing installs update themselves.

Note on the Windows build: the Windows installer in this release was built from commit 6b2898f (Windows shell fixes + the computer-use latency work). It does not include the box-agent streaming/settle fixes listed above, which landed a few minutes later in 20de31e and are present in the macOS build only — they reach Windows in the next release. Everything else is identical across platforms.

Assets 12
Loading