Question 1

Does letting Jev act actually save money, or just driver steps?

Accepted Answer

Driver work, yes; money, not yet. The dual arm makes 23 to 29 percent fewer driver requests than the guarded arm across two same-night samples and runs its own steps at a median 1.4 seconds against 15.7 seconds for an LLM step, but it pays for that in Jev call volume. Task wall time between the arms is not distinguishable on live sites, and dual matched the stock arm's cost only on the fixture heldout so far.

Question 2

Is the zero false completions result just the author's own judge?

Accepted Answer

The interesting part is that the un-guarded arm's failure was accepted by its own LLM judge: told to reach a page, stock deleted an account and recorded it as a success. Guarded and dual recorded zero false completions on every split and run, on predicates checked against observed state rather than self-report, judged on the page and the trajectory, with answers required to quote evidence atoms found in the full page text.

Question 3

What did the author try that did not work?

Accepted Answer

Two things, reported plainly. Delegating bounded subgoals from System 2 to Jev cost more and produced no more completions across three measurements, with the executor reaching 5 of 30 assignments before an audit found software defects behind most failures, now fixed but not re-measured. Declaring the contract to the driver raised model requests 21 percent per task with completions unchanged. Delegation stays available as a tool, off by default.

Question 4

Can I run it on my Mac without it stealing my focus?

Accepted Answer

That is the default. Sessions start in background_only mode, leaving your front app, pointer and keyboard alone, and foreground_permitted only holds the front for 36 to 65 ms and only after 3 seconds of idle. All 15 coexistence cases passed every check, including a conflicting edit and a foreground step beside an idle or moving person. It does need Accessibility granted to whichever process runs it.

Question 5

How do I drive Footwork from the terminal?

Accepted Answer

Through the footwork CLI: footwork start with an app and a requirement begins a session, footwork look shows the menu of what can be done now with ids and the last receipt, footwork s1 --act shows System 1's proposal on the current observation, footwork do click runs an action through the gate, and footwork done has the verifier judge the claim.

Question 6

How does the guard verify a claimed completion?

Accepted Answer

A done is accepted only when the complete noul and every requirement's unmet noul are inside the accept band, and answers must quote evidence atoms found on the full page text. Requirements that name an action are judged from the trajectory once the page has moved on, and a requirement judged met stays met unless a current-state requirement is later contradicted.

Question 7

What does the destructive-action gate do?

Accepted Answer

The gate applies keyword and noul rules on the target, the URL, the focused element for Enter, and evaluate, scoped by the task's authorized_actions. A session defaults to background_only, so the person's front app, pointer and keyboard are left alone, and nothing dispatches on an id from an older observation.

Footwork

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community