name: scoring-and-fingerprint
description: Scoring checks and fingerprint collection for @rely-ai/caliber. Use when working on caliber score output, adding a new scoring check, modifying project fingerprinting, or understanding how project context is collected and scored.

Scoring and Fingerprinting

Scoring (src/scoring/)

The caliber score command runs a fully deterministic quality audit — no LLM, no network calls.

Check categories and point values

See src/scoring/constants.ts. Categories:
- Existence (25 pts) — CLAUDE.md, cursor rules, skills, MCP servers, cross-platform parity
- Quality (25 pts) — build/test commands documented, CLAUDE.md under 120 lines, no vague instructions, no directory tree listings, no contradictions
- Coverage (20 pts) — actual dependencies named, services/MCP referenced
- Accuracy (15 pts) — documented commands exist in package.json, documented paths exist on disk, config freshness
- Freshness & Safety (10 pts) — no secrets, permissions configured
- Bonus (5 pts) — hooks configured, AGENTS.md, OpenSkills format

Adding a new check

1. Edit the relevant category file: src/scoring/checks/{existence,quality,coverage,accuracy,freshness,bonus}.ts
2. Each check function returns Check[]:

import type { Check } from '../index.js';

const checks: Check[] = [];

checks.push({
  id: 'my-check-id',
  label: 'Human-readable description',
  points: 3,
  earned: conditionPasses ? 3 : 0,
  note: conditionPasses ? undefined : 'Why it failed',
});

return checks;

1. Add the point constant to src/scoring/constants.ts
2. computeLocalScore() in src/scoring/index.ts aggregates all checks automatically

Target filtering

Checks are filtered by target agent (claude, cursor, codex) via CURSOR_ONLY_CHECKS, CLAUDE_ONLY_CHECKS, BOTH_ONLY_CHECKS, and CODEX_ONLY_CHECKS sets in src/scoring/constants.ts. Add new check IDs to the appropriate set if platform-specific.

Displaying scores

src/scoring/display.ts exports displayScore(), displayScoreDelta(), displayScoreOneLiner() — all use chalk for formatting.

Fingerprinting (src/fingerprint/)

Collects structured project context before sending to the LLM for config generation.

What gets collected

Key design rule

All detection is 100% LLM-driven. There are NO hardcoded mappings of file extensions to languages or filenames to tools. collectFingerprint() is async and always includes LLM enrichment internally — there is no separate enrichment step.