Question 1

Are those context-reduction numbers from a real workload?

Accepted Answer

They are the project's own measurements, with raw result files in the repository. The 48-run benchmark reported 88 to 97% less tool context, a separate public benchmark used 56.5% fewer input tokens and ran 32.4% faster than single-record parallel calls, and there were 8 of 8 exact fixture runs plus three installed workflow acceptance checks. None of that describes your traffic, and the package itself argues you should keep exact paths, IDs, selectors and arithmetic in native tools.

Question 2

Can it really drive a browser, or is that a demo?

Accepted Answer

It drives a page over CDP against your own signed-in browser or through a named session, and on the project's test shop it pauses before the order button until you approve, then resumes the same tab and runs a verifier against text you specify. Only status: done with exit 0 is success. needs_confirmation, blocked, needs_value and a command that cannot start all end non-zero with the reason on stderr, and --dry-run returns the chosen control without acting.

Question 3

What stops it from deleting my production data?

Accepted Answer

Confirmation, plus an executable allowlist. Jev only chooses among controls the program enumerates, so it cannot invent a selector or type a command, and a destructive goal such as deleting all records ends as needs_confirmation before anything happens. You also get --dry-run, a stop file at ~/.jev-filter/STOP, and the option to park the pointer in the top-left corner. Terminals, password managers and system settings are refused outright.

Question 4

How accurate is survey labelling on my own data?

Accepted Answer

Unknown until you measure it. On 2,000 generated tickets the README reports 100% topic and 95.2% sentiment accuracy against the generator's labels, then says outright that those records are easy by construction and to use --labels on your own set. Uncertain and failed IDs stay in the report rather than being resolved for you, and that is exactly where the real accuracy risk sits, so read the review buckets before trusting the counts.

Question 5

How do I install it, and which platforms are supported?

Accepted Answer

Node.js 22 or newer on macOS or Linux, with Python included in the npm distribution; on Windows use WSL or install Python natively. Install with npm install -g @apixly/jev-filter and run jev-filter doctor, which is offline. Only inference needs a TypeSafe key, exported as TYPESAFE_API_KEY.

Question 6

How do I know a result is incomplete rather than failed?

Accepted Answer

Parse the packet even when the exit code is 2. A result carries selected_ids, review_ids and complete; on exit 2 you get complete false and a review_ids list that needs attention rather than an empty answer. Exact paths, IDs, selectors, calculations and short outputs are meant to stay in native tools, so a reviewed record can always be recovered by its original ID.

Question 7

Does a smaller context actually cost less and run faster?

Accepted Answer

Not necessarily, and the README reports the negative results too. Five of the six whole-operation cells were slower by 0.1 to 18.4%, and the cold API-equivalent estimate ranged from 23.2% lower to 1.1% higher on one cheap-model exec cell. The savings dashboard reports input-equivalent estimates, not invoice savings, and subscription billing remains unknown.

Jev Filter

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community