JevMade Sign in
← Back to experiments

Agent tooling

PiJev

Jev helps Pi choose skills, rank source files and understand failed commands.

Bookmark: PiJev Keep this in your collection.
Leave a noteWhat would you try with this? : PiJev

Only you can see your notes.

Source screenshot of PiJev
SOURCE SCREENSHOTFull screenshot ↗

What it does

The maker found fewer tool calls but no better task completion rate.

How you can use it

You can borrow this idea to help an automated coding tool sort through broken tests. Before writing code, make a list of common software mistakes and simple labels for them, like network trouble or typos. Your developer can then set up an access key to send those errors out for advice.

Keep in mind that external advice is only a suggestion, not a guaranteed fix. Your app will share your code snippets and error logs with an online service to rank them. Your developer must still run real tests to confirm any suggested changes work properly.

Maker-reported (not independently measured by JevMade): Reranking BM25's top-100 on a ~2,000-file django corpus lifted recall@1 of reference-patch files from 0.25 to 0.74 and recall@10 from 0.74 to 0.96, median first gold-file rank 4 to 1 (author-reported) · Tool calls fell 16% on django, 18% on unfamiliar repositories with v4-flash and 25% with v4-pro; 15/20, 11/13 and 12/13 tasks were shorter (author-reported) · The whole retrieval sweep cost $0.35; Jev ranking costs $0.011-0.014 per task (author-reported) · Resolve rate was unchanged: 4 vs 4 on unfamiliar tasks and 15 vs 14 on django, one run per task (author-reported)

Primitives
choice, noul
Platform
TypeScript
Added
Project created
GitHub stars
4 (snapshot, not live)

Keep this for later

Sign in to bookmark experiments, guides and videos, and keep notes only you can see.

Continue to sign in

We’ll bring you back to this listing.