Issue Radar

Personalized GitHub issue discovery, built on GitHub's hybrid semantic issue search (/search/issues?search_type=hybrid, GA April 2026).

Given your GitHub profile (languages, merged PRs), it generates semantic queries, searches all public open+unassigned issues, filters noise (bots, tiny repos, stale issues), and ranks what's left by signals like good first issue labels, freshness, and repo popularity.

Run it

cd ~/projects/issue-radar

# one time: create venv + install deps
/opt/homebrew/bin/python3.14 -m venv .venv
.venv/bin/pip install fastapi "uvicorn[standard]" jinja2 httpx

# start the web app (uses your `gh` CLI login automatically)
.venv/bin/uvicorn main:app --port 8000

Open http://localhost:8000 — first load takes ~15s while it builds the feed. Refresh button (or ?refresh=1) rebuilds; otherwise results cache for 6h.

CLI spike (no server)

python3 spike.py   # works with system python3, uses `gh api`

Auth

Resolution order: $GITHUB_TOKEN env var → gh auth token.

Any GitHub user can run this against their own profile — it always reads the profile of whoever the token belongs to.

Files

file	what
spike.py	zero-dependency CLI prototype (validation)
engine.py	matching engine: profile → queries → search → score
main.py	FastAPI app, in-memory 6h cache
templates/feed.html	the feed UI

Tuning

In engine.py: MIN_STARS (repo quality floor), MAX_AGE_DAYS, PER_PAGE, MAX_QUERIES, and the score() weights.

Making it multi-user (deploy path)

Right now it's single-user/local. To let anyone log in:

1. Register a GitHub OAuth App (github.com/settings/developers), callback URL <host>/auth/callback
2. Add OAuth login routes (Authorization URL → callback → token in session)
3. Replace engine.get_token() with the session user's token
4. Per-user rate limits (10 semantic req/min) mean it scales naturally; add SQLite caching so repeat visits don't re-search

Gotchas learned (don't rediscover these)

• stars: is NOT an issue-search qualifier — hybrid search degrades it into a text token. Filter stars client-side (GraphQL batch fetch).
• PR titles need repo-specific identifiers stripped before they become queries, or they poison the embedding (0 results).
• Over-generic queries (single common word) return 10k+ results = noise.
• Semantic/hybrid search is rate-limited to 10 req/min per token — batch nightly, cache aggressively, never search live per page view.