Measure an AI Coding Harness from L0 to L4 with Harness Score

Описание и применение

Пошаговый процесс повышения уровня зрелости репозитория AI‑кодинга от L0 до L4 с помощью Harness Score: создание CLI, добавление контекста, правил, гигиены, тестов, CI и хуков, при этом каждая итерация проверяется сканером и фиксируется в репозитории.

Задача

AI coding assistants can edit a repository quickly, but speed does not indicate whether the repository can catch a bad edit.

Результат

The repository progresses from L0 to L4, gaining durable context, scoped guidance, hygiene, sensors, CI, and runtime guardrails, resulting in a deterministic maturity score that serves as a diagnostic and gate rather than a proof of correctness.

Рабочий процесс

• Create a disposable copy of the harness-score-tutorial template
• Run npx --yes harness-score . to establish baseline L0 and record score
• Add a minimal functional Meeting Cost CLI and commit
• Add a substantive AGENTS.md for durable context and re-run scan
• Add scoped rules, reusable skills, .gitignore, lockfile for hygiene and re-run scan
• Add tests, formatter, linter, type checks, and GitHub Actions workflow for sensors and re-run scan
• Add Cursor hooks for gate and feedback, tests for hook payloads, and a CI job that runs npx --yes harness-score . --min-level 4
• After each stage, review git diff --stat, run local checks, re-run harness-score, and inspect JSON output for failed checks

Доказательства

AI coding assistants can edit a repository quickly, but speed does not tell you whether the repository can catch a bad edit.

Источник

The current published package is 1.7.5 and the current template clone is expected to start at L0, because it contains the tutorial and license but not the application harness.

Источник

Run npx --yes harness-score . and record the initial level and score.

Источник

Add a substantive root AGENTS.md. It should describe the real files, commands, domain invariants, error handling, and security boundaries.

Источник

Add tests for valid and invalid calculations, a formatter, a linter, strict type checking, and a GitHub Actions workflow.

Источник

The scanner measures filesystem and configuration facts. It does not establish that the Meeting Cost CLI is commercially useful, that tests are comprehensive, that rules are correct, or that a team actually reviews pull requests.

Источник