Question 1

What is Jev Issue Radar?

Accepted Answer

A read-only triage dashboard that finds likely duplicate GitHub issues with the evidence side by side. It retrieves a small candidate set, asks Jev to classify each pair and shows selected passages; the model selects a relationship and evidence IDs, and the maintainer decides.

Question 2

How does the duplicate-detection pipeline work?

Accepted Answer

Issue URL → retrieve up to 300 recently updated items → exclude PRs and the source issue → rank up to 5 candidates by title-weighted lexical overlap → Jev answers four Choice questions → validate choices and show the original passages. Nothing is closed or labelled.

Question 3

Can I try it without an API key?

Accepted Answer

Yes — the public sample on GitHub Pages needs no signup, installation or key. It contains four synthetic reports with clearly labelled, hand-authored decisions, and the README describes it as a demo of the flow rather than a quality claim.

Question 4

What does the local app need?

Accepted Answer

Node.js 22+ with built-in HTTP and fetch, a server-side OpenRouter key, and read-only GitHub REST access. There are no production dependencies or build step; the browser talks to a loopback Node server that holds the key.

Question 5

Does it ever close issues automatically?

Accepted Answer

No. It never closes issues, adds labels or posts comments — stated as a design constraint. A duplicate is displayed only under strict rules (both evidence IDs exist, compatible reason, no truncation, confidence at least 0.8), and that rule is a product choice, not a calibrated accuracy guarantee.

Question 6

How accurate is the duplicate detection?

Accepted Answer

Not established. The README says the integration works but duplicate-detection quality on real repositories has not been established. In the initial four-pair synthetic smoke test two matched the author's expected labels; in a later three-pair check 0/3 canonical issues reached the top five, which the docs attribute to retrieval.

Question 7

How is it tested and what state is it in?

Accepted Answer

Tests use node:test with optional Playwright, separating application correctness from model quality. The project ships docs/VALIDATION.md, docs/PUBLIC-CASES.md and a CI workflow, and marks itself an experimental MVP.

Jev Issue Radar

Highlights

Questions about this entry

From the Jev guides

From the Jev AI model community