Question 1

How does the two-stage Jev check actually work?

Accepted Answer

It filters by description with Choice ranking, checks usefulness with a separate Noul per group keeping at most three candidates, then re-checks body snippets with Choice plus a fit Noul. Jev returns typed judgments only; the host must read the full entry and verify availability before use.

Question 2

Can I treat the 0.80 fit gate as 80% accuracy?

Accepted Answer

The 0.80 fit and 0.65 confidence gates are documented as exploratory thresholds, explicitly not calibrated correctness rates. Treat them as triage gates and always have the host verify the recommended entry.

Question 3

What key and privacy exposure does Jev mode involve?

Accepted Answer

Local mode needs no key and gives keyword candidates only; Jev mode needs a TypeSafe key and pins a fixed model version. Explicitly named skills resolve locally with zero API calls, and reports plus cache files use restrictive permissions.

Question 4

What should my wrapper handle besides a suggestion?

Accepted Answer

Expect none, uncertain, ambiguous, local_only, or incomplete alongside suggested and explicit_selection. Exit zero covers completed flows including no-match; nonzero codes cover input, setup, and incomplete Jev runs. Integrate on the JSON status field.

Question 5

How do I install it as a Codex Skill?

Accepted Answer

Run the documented copy snippet in the clone, which copies the six runtime files into ~/.codex/skills/jev-skill-suggester and errors if the target already exists so it cannot overwrite an install. Then open a new session and invoke $jev-skill-suggester.

Question 6

What did the first live test measure?

Accepted Answer

On 16 pre-labeled tasks over 24 entries and 23 eligible candidates, 15 of 16 produced the expected complete result, 12 of 12 matching tasks recommended the preset skill, 3 of 4 no-match tasks returned none, and 0 of 16 gave a wrong final recommendation. It used 45 HTTP attempts and 230,067 input tokens at a median 5.05 seconds per task, an estimated $0.009662814.

Question 7

How do I limit scope and budget?

Accepted Answer

Use --allow and --exclude to restrict candidates, --require for a user-named skill, --max-calls for 1 to 64 HTTP attempts per run with a default of 12 including retries, and --cache-dir for a 24-hour response cache keyed by task, model, question and skill content hash.

Jev Skill Suggester

Highlights

Questions about this entry

More Jev resources like this

From the Jev community