Back to search
shipgate
MarketplaceWorks with Claude CodeGitHub
Gated multi-agent delivery pipeline for Claude Code: test-first, locked tests, cross-model review, single testing queue. Jira or GitHub Issues.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Checked today
Badge for your README
Markdown
[](/listings/iremir0--shipgate)HTML
<a href="/listings/iremir0--shipgate"><img src="/api/listings/iremir0--shipgate/badge.svg" alt="findsafeskills safety rating" /></a>Install
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 21 lines
- {
- "name": "shipgate",
- "owner": { "name": "Emir Tüfekci", "url": "https://github.com/Iremir0" },
- "metadata": { "description": "Gated multi-agent delivery pipeline for Claude Code" },
- "plugins": [
- {
- "name": "shipgate",
- "source": "./plugins/shipgate",
- "description": "Ticket-to-release pipeline: test-first, locked tests, cross-model review, single testing queue. Jira or GitHub Issues.",
- "version": "0.1.0",
- "category": "development"
- },
- {
- "name": "shipgate-hooks",
- "source": "./plugins/shipgate-hooks",
- "description": "Optional Stop hook that runs your quick checks on changed files before Claude ends a turn.",
- "version": "0.1.0",
- "category": "development"
- }
- ]
- }
README.md · README · 139 lines
- <p align="center">
- <img src="assets/banner.svg" alt="shipgate: To Do → In Progress → Code Review → Testing → Ready for Release" width="100%">
- </p>
- <p align="center">
- <a href="LICENSE"><img alt="MIT" src="https://img.shields.io/badge/license-MIT-d97757"></a>
- <img alt="version" src="https://img.shields.io/badge/version-0.1.0-2a1d16">
- <img alt="Claude Code plugin" src="https://img.shields.io/badge/Claude%20Code-plugin-d97757">
- <img alt="trackers" src="https://img.shields.io/badge/trackers-GitHub%20Issues%20%7C%20Jira-2a1d16">
- <a href="https://github.com/Iremir0/shipgate/actions/workflows/ci.yml"><img alt="CI" src="https://github.com/Iremir0/shipgate/actions/workflows/ci.yml/badge.svg"></a>
- </p>
- <p align="center"><b>A gated, multi-agent delivery pipeline for <a href="https://docs.claude.com/en/docs/claude-code">Claude Code</a>.</b><br>
- You write tickets and release. Agents write tests, code and reviews. Gates decide what ships.</p>
- <p align="center"><a href="README.tr.md">Türkçe</a> · <a href="docs/how-it-works.md">How it works</a> · <a href="https://github.com/Iremir0/shipgate-demo">Demo repo</a> · <a href="docs/case-study.md">Case study</a></p>
- ---
- ## Why
- Letting one AI session "just build the feature" works until it doesn't: it edits the test to make it pass, reviews
- its own blind spots, and calls something done that was never run against the real services.
- This plugin is the process I use on a production e-commerce codebase instead. The main Claude Code session becomes a
- **manager that is not allowed to write code**. It moves each ticket across a kanban board and hands the work to
- specialised agents, and a ticket only moves forward when a gate is passed with evidence.
- ## The pipeline
- ```mermaid
- flowchart LR
- A[To Do] --> B["In Progress<br/><sub>test-author → dev</sub>"]
- B --> C["Code Review<br/><sub>other model + security</sub>"]
- C -- reject --> B
- C --> D["Testing<br/><sub>single queue, merged with main</sub>"]
- D -- fail --> B
- D --> E[Ready for Release]
- E -- you --> F[Done]
- ```
- | Gate | Rule | Why |
- |---|---|---|
- | 🔒 **Locked tests** | Failing tests are written first, in their own commit. The dev agent may not touch them; the reviewer diffs them and auto-rejects any change. | The agent can't "fix" the test instead of the code. |
- | 🔀 **Cross-model review** | Sonnet's code is reviewed by Opus and vice versa. A forced same-model review is stamped ⚠️ on the PR. | A model approves its own mistakes. |
- | 🛡️ **Risk routing** | Money, payments, auth, migrations → stronger dev, mandatory test-first, security review (Critical/High = reject). | Spend the expensive checks where mistakes are expensive. |
- | 🧪 **Single testing queue** | One ticket at a time, merged with the latest `main`, every relevant suite incl. e2e against real services. | Shared DBs and ports; two green branches can be red together. |
- | ⛔ **BLOCKED ≠ PASS** | A suite that couldn't run never counts as passed. | No "tests skipped, looks good". |
- | 🔁 **Loop limit** | 3 rejects + fails → `blocked`, human looks. | Stops agents from grinding on a bad ticket. |
- | 🙋 **Humans release** | The pipeline stops at *Ready for Release* with a release note and a pipeline log. | You stay accountable for what ships. |
- More detail and the reasoning behind each rule: **[docs/how-it-works.md](docs/how-it-works.md)**.
- ## Try it in 5 minutes
- The **[demo repo](https://github.com/Iremir0/shipgate-demo)** is a tiny cart calculator with three seeded
- issues: a copy change, a risky money feature and a planted bug.
- ```bash
- gh repo create my-pipeline-demo --template Iremir0/shipgate-demo --private --clone
- cd my-pipeline-demo && npm ci && npm run seed
- claude # trust the folder, accept the plugin install
- > /pipeline
- ```
- Watch the issues move across `status:*` labels, with plans, review decisions and release notes as comments and one
- merged PR per issue. The first real run (3 issues, 3 merged PRs, and the 8 plugin problems it surfaced) is
- written up in **[docs/example-run.md](docs/example-run.md)**.
- ## Use it in your repo
- ```text
- /plugin marketplace add Iremir0/shipgate
- /plugin install shipgate@shipgate
- /plugin install shipgate-hooks@shipgate # optional Stop hook
- /pipeline-init # detects your stack, writes .claude/pipeline.json, sets up the tracker
- /new-ticket add refund requests
- /pipeline # work the whole board
- /pipeline #12 # or one ticket
- ```
- | Command / agent | What it does |
- |---|---|
- | `/pipeline [key]` | The manager. Board mode or a single ticket. |
- | `/new-ticket <idea>` | Drafts a ticket with testable Given/When/Then criteria, test plan, out-of-scope and security notes; creates it after you approve. |
- | `/pipeline-init` | One-time setup per repo. |
- | `test-author`, `dev-standard`, `dev-critical`, `code-reviewer`, `security-reviewer`, `test-runner` | The agents ([agents/](agents)). |
- | `shipgate-hooks` (optional plugin) | A Stop hook that runs your quick checks on changed files before Claude may end a turn. |
- Configuration reference: **[docs/configuration.md](docs/configuration.md)** · Trackers: **[docs/trackers.md](docs/trackers.md)**
- ## In production
- I run this pipeline on a production e-commerce monorepo (Medusa 2 + Next.js 16 + Expo, Jira kanban) where prices follow
- the live gold price, so most work touches money, stock or payments. First three weeks:
- | 30 | 76 % | 81.5 % | 3 |
- |:---:|:---:|:---:|:---:|
- | tickets delivered | approved on first cross-model review | passed first test run against `main` + real services | test runs reported **BLOCKED** instead of a fake pass |
- One ticket in four was sent back by the reviewer and one in five failed integration before any human saw it.
- The full story, including what went wrong and what changed because of it: **[docs/case-study.md](docs/case-study.md)**.
- ## What shipgate runs and fetches
- shipgate has no server, no telemetry and no dependencies. Everything happens on your machine, through tools you
- already authorised:
- - **Stop hook**, only if you install the optional `shipgate-hooks` plugin
- (`plugins/shipgate-hooks/hooks/verify-on-stop.mjs`, plain Node): runs `git status` and the commands *you* list under
- `verifyOnStop` in `.claude/pipeline.json`; writes a fingerprint cache to `.claude/.cache/`.
- - **`/pipeline` and the agents** run `git` (worktrees, branches, push of ticket branches), your project's commands from
- `.claude/pipeline.json`, and your tracker:
- - GitHub Issues: the `gh` CLI with your own login (issues, labels, PRs).
- - Jira: the Atlassian MCP server you connect yourself.
- - **`/pipeline-init`** reads your project files to propose a config and creates GitHub labels with `gh`.
- - Nothing is downloaded at runtime. The JSON schema URL in `pipeline.json` is only used by your editor for autocomplete.
- ## FAQ
- **Isn't this expensive?** Cheap work stays cheap: copy and small changes go to Sonnet and may skip test-first. Opus is
- spent on risky code and on reviewing Sonnet. The pipeline log on every ticket tells you where the tokens went.
- **Can it merge to my main branch by itself?** By default yes, after review *and* testing pass; it never pushes to
- `main` directly. Set `"autoMerge": false` to stop at an approved, tested PR.
- **What does it ask me?** Only real decisions: a new module or data model, where data lives, a new dependency, a
- customer-visible rule with several reasonable answers. Everything else it decides and records on the ticket.
- **Does it work without Docker / integration tests?** Yes. `commands.suites` is optional; the demo has none.
- ## Roadmap
- - Linear and GitHub Projects (v2) status fields as trackers
- - Cost per ticket in the pipeline log
- - A `/pipeline-report` command that charts reject/fail rates per model over time
- ## License
- MIT © Emir Tüfekci
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.