Tracer

Overview

Searching `good first issue` returns thousands of results that are stale, already claimed, or far harder than the label suggests. Tracer reads the repository, the issue and your skill profile, then gives each opportunity a Contribution Fit Score, the reasoning behind it, and a plain verdict on whether to take it. The scoring engine is deterministic; the AI layer is optional and the product is built to run with it switched off.

What I shipped

• Deterministic scoring engine over eight weighted dimensions, from skill match to whether somebody else is already working on the issue. Weights live in one config file that no analysis module imports, so the model can be changed without touching the analysis.
• Verdict guardrails that can only push a verdict down, never up: a high score with an unclear scope or an issue that is not genuinely free caps at 'possible' rather than 'recommended', so the verdict is not a pure function of the score.
• The AI layer is optional by design. The product works with AI_PROVIDER=none, and anything the model inferred is marked as inference rather than observed data.
• Published the methodology in docs/scoring.md, every weight and rule, on the grounds that a recommendation engine nobody can inspect is not trustworthy.
• Two background jobs outside the request cycle: discovery, which searches and scores candidates, and a refresh that re-checks saved issues for closure, assignment or a new pull request.
• Open source under MIT with a contributing guide, a code of conduct, issue templates, and CI running lint, typecheck, tests and the build on every pull request.

Stack

• Next.js 15
• TypeScript
• PostgreSQL
• Drizzle ORM
• Auth.js
• GitHub GraphQL
• Vitest