gabe4coding’s plain turns English instructions into browser tests. Jev selects from described controls and judges claims about the page. Playwright executes the actions. Desktop and mobile adapters extend the same pattern.

Give the judge the missing evidence

The framework reads an accessibility tree: the page’s text, control names, roles, and state. It adds rendered geometry for spatial questions, rather than sending screenshots to Jev.

  • Choice questions select a control or none; yes/no questions judge test claims.
  • The code passes claims at probability 0.9 or higher, fails them at 0.1 or lower, and leaves the middle inconclusive.
  • Its claim report describes a false pass when an editor contained text that the original snapshot omitted.
  • Adding explicit field text, [empty], and [readonly] marks removed the reported false passes and false fails in that comparison.

Those claim results are the maintainer’s experiments, not independently reproduced here. Better evidence changes what the model can judge; a valid answer type cannot recover missing page state.

Measure the whole workflow

The repository’s corrected comparison covers 108 trials across six synthetic browser workflows, three main models, and both tool stacks. An offline recount of its saved records confirms the reported aggregates:

  • With Luna, plain’s mean API cost was 6.5% higher, while elapsed time was 52.1% lower.
  • With Terra, mean API cost was 23.8% lower and elapsed time 25.5% lower.
  • With Astra, mean API cost was 10.3% lower and elapsed time 51.2% lower.

The maintainer’s cost intervals include no savings for Luna and Astra. Costs include both main-agent and Jev calls; timings include provider and browser waits. These compare complete tool stacks, not Jev inference alone.

An earlier comparison had a broken baseline snapshot-file reader. The project discloses it and publishes a fresh corrected comparison. One strict baseline failure involved a duplicate correct save after a fixture reset, rather than editing the wrong account.

Record the result before asking again

  • Judge mode: Jev selects targets and judges claims, recording successful runs in lock files.
  • No-judge mode: recorded locators replay, and claims require the recorded passing state to match.
  • Auto-healing: stored targets replay where possible; Jev repairs misses and still judges claims.

Review changed lock files before committing them. A replay can preserve an initially mistaken judgment, and live page content still reaches the hosted model when it is consulted.

Caveat: source inspected and saved trial records recounted; no live browser/model run or security audit. Typed output is not a correct decision.