Skip to content

0.139.0: Computer-use studies can run on Modal desktops

Choose a tag to compare

@danielgwilson danielgwilson released this 19 Oct 22:08
· 103 commits to main since this release
8fa2f00

humanish 0.139.0 can run a computer-use study's desktops on Modal. Set
execution.target: modal-desktop on an app-url study and each participant gets its own Modal
sandbox running the same desktop the E2B route uses: Xvfb, XFCE, xdotool input, screenshots, a
live noVNC stream, Chromium and recordings. The same study file runs on E2B or Modal
with only the target changed. Clone and local-tree subjects, shared worlds, E2B templates and
camera or speech media stay on E2B.

Added

  • A computer-use study on an app-url subject can run its participants' desktops on Modal:
    execution.target: modal-desktop. Each participant gets its own Modal sandbox running the
    desktop the E2B route uses: Xvfb at the participant's screen size, an XFCE session, xdotool
    input, scrot screenshots and a noVNC live stream through a Modal tunnel, so the computer-use
    loop, the screen and geometry checks and recordings work as on E2B. The browser
    is Chromium. The run needs the Modal token pair, as modal-terminal does, and allows Modal's
    24-hour lifetime; it is journaled and reclaimed by id, records modal-desktop in the bundle and
    is priced at Modal's sandbox rate. Clone, local-tree, desktop-cli and shared-world subjects,
    E2B desktop templates, camera and speech media, and Chrome or Firefox stay on
    e2b-desktop (#1788).

Checked live

  • The same drawDB app-url study (a public open-source diagram editor), one participant with a 5-minute session, ran on E2B and on Modal with only execution.target changed. Both participants added and named two tables and reported where they hesitated. Both runs verified with 16 checks, and both browsers measured the same window bounds. The Modal desktop cost $0.0086 for 1.09 minutes; the E2B desktop $0.011 for 1.25.
  • Two participants ran on two Modal desktops at once. Both reached the goal, and both sandboxes were released by id.
  • A Modal desktop with recording: { audio: true } kept a complete 57-second video with microphone and speaker sources.
  • A study on the packed 0.139.0 tarball ran a participant on a Modal desktop before the tag.

The release changes nothing a participant sees or does, and nothing the analysis reads, on E2B or local studies, so the efficacy benchmark was not rerun; the 0.133.0 result stands.