Small decisions. Interesting possibilities.Submit contentSubmit

Touchpress mobile end-to-end testing

my mobile e2e testing library powered by @agent_device and AI sdk now supports Jev from @typesafeai Same test, same loop: Jev: 29s, $0.003 Haiku 4.5: 37s, $0.07

Added to Jevfast

How it uses Jev

Touchpress can configure TypeSafe Jev as its evaluation model for `device.act`. At each step, the test captures the screen, lists offered moves, and asks which move should come next; Jev returns a typed Choice rather than writing text.

What Jev decides

Illustrative Jev requestExample answers · not a recorded Jev response · Source ↗
Question 1 · next
YOUR APP
INSTRUCTION

Choose the offered move that advances the task or correctly finishes it.

STATE

Task and current mobile screen with no offered element or fill moves; only fixed verdict/navigation moves remain.

JEV · CHOICE
  1. pass
  2. fail
  3. incomplete
  4. wait
  5. back
  6. scroll_up
  7. scroll_down
  8. scroll_left
  9. scroll_right

App workflow

  1. Configure the test target

    Set the app identifier, ready condition, platform, and simulator or emulator in Playwright configuration.

  2. Choose an evaluation model

    Configure the AI SDK evaluation model as TypeSafe Jev for `device.act`.

  3. Drive the app

    Call `device.act` with an instruction; the model chooses among moves available on screen until the instruction is satisfied or it cannot proceed.

  4. Assert the outcome

    Use Playwright locators and expectations to verify the resulting UI state.

Why it is interesting

It brings typed decision-model control into ordinary Playwright mobile tests, so a test can describe an action in English and continue with normal locator assertions.