ai-shipping — плагин для Claude Code: функции и установка
Плагин ai-shipping для Claude Code готовит AI-сгенерированный код к проверке и выпуску. Команда ship-check собирает документы, ревью и карту тестового покрытия в один процесс.
MIT ★ 741 обновлён 2026-11-14
/plugin marketplace add phuryn/product-engineering
/plugin install ai-shipping@product-engineeringЧто делает ai-shipping
Плагин описывает приложение, проверяет изменения и помогает провести их через выпускной шлюз. ship-check запускает весь процесс: готовит документы и контекст для агента, проводит проверки корректности, безопасности и производительности, собирает независимое ревью и карту тестового покрытия.
Отдельные скиллы запускаются по запросу или по типу работы. Статические аудиты принимают диапазон Git-коммитов или незакоммиченное рабочее дерево и проверяют затронутые изменения. В режиме check скиллы сверяют документацию с кодом, но не записывают файлы. Если ship-check собирается перезаписать существующие документы или карту покрытия, он сначала перечисляет изменения и спрашивает разрешение.
Для установки нужны обе команды: сначала добавьте маркетплейс, затем установите плагин. Ревьюер по умолчанию использует Grok CLI, если он установлен; иначе процесс пробует свежую сессию Codex, затем Claude Code.
Какие задачи решает
- подготовить документы приложения для ревью
- сверить описанное поведение системы с кодом
- проверить изменения на ошибки корректности, безопасности и производительности
- составить карту тестового покрытия
- провести независимое ревью перед выпуском
Примеры запросов агенту
- Запусти полный
ship-checkдля этого репозитория. - Сверь документацию приложения с кодом в режиме проверки.
- Проведи независимое ревью изменений перед выпуском.
Как установить ai-shipping
Через маркетплейс плагинов
/plugin marketplace add phuryn/product-engineering/plugin install ai-shipping@product-engineeringПо шагам, с командами для каждой платформы: Плагины Claude Code: маркетплейсы, установка и как собрать свой
Вопросы об ai-shipping
Как установить ai-shipping в Claude Code?
Нет: команды должны быть записаны точно: сначала /plugin marketplace add phuryn/product-engineering, затем /plugin install ai-shipping@product-engineering.
Что делает ai-shipping?
Плагин документирует приложение, проверяет код и готовит проверки перед выпуском. Полный процесс запускает ship-check, а отдельные этапы доступны как скиллы.
Что происходит, если ship-check собирается перезаписать документы?
Он перечисляет файлы, которые собирается изменить, и запрашивает разрешение. В режиме check сверка документации с кодом проходит без записи файлов.
Исходный текст
ai-shipping · MIT
# ai-shipping — ship AI-built apps
For builders, PMs and founders accountable for AI-built code. Document a vibe-coded app, review it for correctness, security and performance defects, have a fresh reviewer that never saw the build clear every push that deploys and every release, prove it with a test harness that fakes the right side, and release it through a gate the owner says go to.
It is where the AI Shipping Kit from [pm-skills](https://github.com/phuryn/pm-skills) moved: its three skills and the five workflows that are commands there, plus `release-gate` and `test-harness`. `pm-ai-shipping` in pm-skills now only points here, so once this pack is installed, turn it off.
## Overview
AI agents write code fast but leave no record of *intent* — what the system should do, who may do what, where the secrets live. Without that record, no human and no auditing agent can tell whether the code is safe to ship. This kit restores reviewability: it documents the system, then audits the gap between what the docs say and what the code does — the class of bug generic scanners miss because they have no model of intent.
Start the whole sequence yourself with the ship check: `/ai-shipping:ship-check` in Claude Code, `$ai-shipping:ship-check` in Codex. Every stage of it is also a skill of its own that starts when you ask for that stage ("document the app", "what should we test", "run a security audit").
Folders and file names are defaults. The kit follows the repo's own setup — its agent instructions, its docs folder, and the documents and test-coverage file it already has — and falls back to `documentation/`, `documentation/tests.md` and `reports/` only where the repo has none, saying so in its output.
The two static audits take a git range (`<base>..<head>`, such as a release's unreleased commits) or `changes` (the uncommitted working tree) and then audit only the change: the changed files plus what the change makes newly reachable or newly wrong elsewhere. A release normally audits only its changes; audit the whole repository for the first release, or after a large refactor.
For a repo that keeps its docs and coverage map by hand as intent, `document-app` and `derive-tests` in `check` mode are the release-time form ("check the docs against the code"): they write no file, list each disagreement with `file:line` on both sides, leave the owner to decide whether the code or the doc is wrong, and end with a one-line summary a release checklist can cite.
## Install
```bash
claude plugin marketplace add phuryn/product-engineering
claude plugin install ai-shipping@product-engineering
```
Codex and other agents: see the collection's root `README.md`. Any skill can be started by name: `/ai-shipping:<skill>` in Claude Code, `$ai-shipping:<skill>` in Codex. All but `ship-check` also start on their own when the work matches.
## Skills (11)
**You start it:**
- **ship-check** — Turn a vibe-coded repo into a reviewer-ready shipping packet: document, wire agent context, run correctness, security and performance reviews, add an independent unsteered pass by a fresh reviewer that never saw the build (the reviewer the repository's instructions name, when it can run the way `independent-review` requires, and when it can't, their fallback runs if it can, otherwise the wrapper's reviewer runs without asking; if no other family can run, a fresh session of the builder's family, with the actual reviewer and why recorded; with none named, through the `independent-review` wrapper it carries: Grok when its CLI is installed, so another model family; without it a fresh read-only Codex session, then a fresh read-only Claude Code session; with none of the three CLIs, a fresh read-only session of your own agent, given the prompt the wrapper writes), map test coverage, and compile the results. Only when you start it, because it writes docs, the agent instructions and the coverage map, and takes a long session; where docs already say what the product should do, or whoever runs the gate keeps the coverage map, it lists what it would overwrite and asks first.
**They start on their own:**
- **shipping-artifacts** — The durable documentation set that makes an AI-built app reviewable: a core every app needs (architecture, user/permission flows, permissions, variables/secrets, test-coverage map) plus conditional docs added only when they apply (emails, cron, SEO, embedded agents/automation). Defines what each doc must capture and how a reviewer uses it.
- **code-review** — The top-level review skill. Correctness is its core; **performance and security are optional sub-cases** of the same engine. Anchors on agreements between participants across a boundary — the defects that are invisible file-by-file because each side looks reasonable alone — forces a violating execution, and refutes every candidate before reporting.
- **independent-review** — The unsteered review loop before anything ships: a fresh reviewer that never saw the build reads the unreleased diff and the working tree, round after round, until a round finds nothing the change introduced above Low. By default the reviewer is the Grok CLI when it's installed, so another model family; without it a fresh read-only Codex session, or else a fresh read-only Claude Code session. Your agent instructions can name another, or another order, that runs the way the skill requires. How to scope a round, write its brief, triage what it finds and when to stop, through a wrapper that writes the subject to disk and fails a round that came back too short to have run.
- **release-gate** — How a change ships: the three moments and their bars (a stage done, a push that deploys, a promotion), the owner's preview, the preconditions, go/no-go, promote, and proving the build serving is the one reviewed. Learned on Next.js + Supabase + Clerk + Netlify, principle first.
- **test-harness** — How to test an AI-built app so a green gate means something: what each harness fakes, the script contract, the authenticated mile, the rules that came from bugs, what to do when a fix doesn't land, and the coverage map written by whoever ran the gate.
- **intended-vs-implemented** — The method for finding the gap between what a system is documented to do and what the code actually does, with cited evidence on both sides and without hand-wavy findings.
- **document-app** — Reverse-engineer a codebase into the system documents reviewers and auditors need — a core set (architecture, flows, permissions, variables) plus conditional docs (emails, cron, SEO, automation) when they apply. In `check` mode it writes nothing and reports where the existing docs and the code disagree.
- **derive-tests** — Turn documented intent into a test-coverage map: inventory the tests that exist today, separate them from proposed tests and unverified gaps, mark each by what it depends on, and recommend a gate that follows the repo's stated gate policy (a green-before-merge CI gate where it states none). In `check` mode it writes nothing and reports where the coverage map is stale.
- **security-audit-static** — Static security audit: map trust boundaries, cross-reference documented intent, self-refute every finding, and report only evidence-backed risks. Given `<base>..<head>` or `changes`, it audits only a change. Starts on its own only when a security audit is asked for: `ship-check` runs it, and `release-gate` proposes it and never runs it unasked.
- **performance-audit-static** — Static performance audit: find N+1 queries and request waterfalls, over-fetching, missing indexes, and caching opportunities, ranked by effort and impact. Given `<base>..<head>` or `changes`, it audits only a change. Starts on its own only when a performance audit is asked for: `ship-check` runs it, and `release-gate` proposes it and never runs it unasked.
## Author
Paweł Huryn — [The Product Compass Newsletter](https://www.productcompass.pm)
## License
MIT
© Paweł Huryn · MIT · phuryn/product-engineering