Validate the candidate, not just the answer type
The adapter rejects unknown candidate IDs, non-finite confidence and out-of-range probabilities before returning a ProviderChoice.
Inspect the supporting source ↗REPOSITORY ANALYSIS
cua
Cua’s preview jev-use example pairs Driver observation and execution with bounded Jev browser-action choices.
E2 · source reviewedPublic material or source evidence was reviewed. A fixed-version source review requires checking a pinned revision; this is not a reproduction, performance endorsement, or safety guarantee.
- Stars
- 26,156
- Forks
- 1,813
- License
- MIT
- Last commit
- 2026-09-24
What Jev does here
Reads DOM or supported visual-region descriptions and returns a supplied candidate action ID.
Provides Python and TypeScript loops with separate offline and live checks.
The Jev integration is a public-preview example, not the entire Cua framework or CUA-S1. Offline fixture checks do not establish live model performance. README and integration source reviewed at a fixed commit; not independently run or benchmarked by this site.
What task does this project fit?
Build a browser-action loop whose success is verified outside the decision model.
Material analysed：
From input to action
- Driver supplies a fresh DOM snapshot and, when supported, visual-region descriptions.
- The adapter asks Jev to choose one supplied candidate ID; the runner resolves that ID to a concrete action.
- After acting, the runner observes again and checks the fixture’s independent /state result.
Implementation details that matter
A mock proof is a different test from a live run
The recipe provides separate deterministic and live modes; a submitted event alone is insufficient without the independently read fixture state.
Inspect the supporting source ↗When to choose it
Choose this recipe when you want Driver execution plus an inspectable bounded chooser. It is an example integration, not evidence that the entire Cua framework uses Jev.
How to inspect it yourself
Inspect choose_with_typesafe in jev_adapter.py and the README’s mock-proof section. Trace one supplied candidate ID through the choice validator and then through the fixture state check; verify that an unknown ID cannot reach execution.
Limits of this review
- No Driver or live Jev run was performed for this review.
- A successful local fixture does not measure reliability on arbitrary websites.