GitHub Issue Scorer

Your backlog has four hundred open issues and you have Tuesday afternoon. Which three actually matter?

Point this agent at a repository and it reads each open issue — the body and the whole comment thread — then fills a live board as it goes. Every card tells you how urgent the issue is, how the reporter feels about it, and how badly users are actually impacted. The frustrated thread where someone quietly posted a workaround, or mentioned they're evaluating a competitor, stops being invisible.

What you get

• A board that fills in as it thinks — cards appear immediately as placeholders and resolve one by one, so you can start reading the first result while the rest are still being scored.
• Scores you can calibrate your trust against — every priority and sentiment carries a confidence figure. Anything under 60% is highlighted, so a shaky judgement invites a second look instead of sitting there looking authoritative. Hover it for the full breakdown.
• A real ranking, not three buckets — an impact score orders issues within each priority, so the top of a HIGH run is genuinely where to start. Each card says what its impact level means in plain words.
• The things a summary usually loses — competitor comparisons and user-discovered workarounds get pulled out of the thread. And when the agent detects one it couldn't cleanly quote, the card says so rather than showing nothing — a nudge to go read that thread yourself.

Usage

Open the dashboard, type a repository, choose how many issues to score, and hit the button.

Cards appear as placeholders the moment a run starts, then resolve one at a time.

Cards sort themselves highest-priority first, then by impact within each priority, so the list arrives in the order you should work through it.

How the scoring works

Two models, each doing the half it is actually good at.

Jev makes the judgements — priority, sentiment, and impact. It is a decision model rather than a chat model: it picks from a fixed set of answers and returns a probability for each, so it cannot invent a priority that doesn't exist or mangle its own output. That is where the confidence figures come from.

An LLM does the writing — the summary, the reasoning behind each score, and pulling competitor names and workarounds out of the comments.

They run at the same time, so you don't wait for both. If Jev is unavailable the agent keeps working and falls back to the LLM's own judgement; you'll just see the confidence and impact figures disappear from the cards.

Limitations

• Read-only. It analyses issues; it never comments, labels, or closes anything on GitHub.
• Open issues only. Closed issues and pull requests are skipped.
• Up to 50 issues per run, and it only sees what the connected GitHub account can see — private repos need access.
• Very long threads get trimmed. Issue bodies and individual comments are truncated, and only the first 120 comments are scored, so a thousand-comment epic will lose some nuance.
• Sessions don't persist. Scores live as long as the page is open; reloading starts fresh.