Six AI subagents that handle the boring repo-maintenance work for you. Drop-in for Claude Code.
English | 简体中文
curl -fsSL https://raw.githubusercontent.com/rockscy/shipcrew/main/install.sh | bashEach agent is a single Markdown file with a focused, opinionated system prompt. Drop into ~/.claude/agents/ and Claude Code will route the right work to the right agent automatically.
| Agent | What it does |
|---|---|
| pr-reviewer | Deep review of a single pull request — line-cited findings, no bikeshedding. |
| issue-triager | Classifies a new issue into one of seven buckets and drafts a reply. |
| release-notes-writer | Turns a tag range into user-facing release notes + a launch blurb. |
| test-failure-investigator | Investigates one failing test, names a root cause, proposes a one-line fix. |
| dep-updater | Reads dependency manifests and produces a triage table (safe / risky / breaking). |
| flake-hunter | Scans CI history, identifies flaky tests, recommends quarantine + tracking issues. |
Most "agent" repos give you a framework to build agents. shipcrew gives you the agents themselves — pre-written, scoped, opinionated.
| Agent frameworks (jcode, etc.) | shipcrew | |
|---|---|---|
| You write | The agent | Nothing — install and use |
| Output style | Free-form | Structured (tables, line citations, fixed sections) |
| Boundaries | Implicit | Explicit "what NOT to do" in every agent |
| Calibration | Confidence not addressed | Each agent emits high/medium/low confidence |
| Composability | Custom integration | Drops into Claude Code's built-in subagent system |
Each agent in the crew is single-purpose, line-cited, and uncertainty-aware. They tell you when they don't know something instead of inventing answers — that's the difference between a useful subagent and an agent that wastes your time.
curl -fsSL https://raw.githubusercontent.com/rockscy/shipcrew/main/install.sh | bashDrops six .md files into ~/.claude/agents/. Existing files of the same name are skipped — never overwritten.
git clone https://github.com/rockscy/shipcrew.git
cp shipcrew/agents/*.md ~/.claude/agents/In Claude Code, ask: "What subagents are available?" You should see pr-reviewer, issue-triager, etc.
Most agents trigger automatically when you describe a matching task. You can also call them by name:
> Review PR #1247 in this repo.
[Claude routes to pr-reviewer]
> Triage this issue: <paste body>
[Claude routes to issue-triager]
> Why is test_checkout_flow_with_coupon failing in CI?
[Claude routes to test-failure-investigator]
> Draft release notes for v1.4.0 (previous tag was v1.3.2).
[Claude routes to release-notes-writer]
You can also explicitly invoke an agent: "Use the dep-updater agent to look at requirements.txt."
- Single-purpose. Each agent does one thing. If you find yourself wanting to make
pr-reviewerdo triage, that means triage needs its own agent. - Line citations. Every finding includes
<file>:<line>. Vague claims are worse than no claims. - Honest uncertainty. Each agent emits confidence: high / medium / low. Wrong-but-confident is the worst output.
- Boundaries first. Each agent explicitly states what it won't do — never approves PRs, never disables tests as a fix, never invents file paths.
- Output formats are concrete. Tables, code blocks, fixed sections. No "let me give you some thoughts" prose.
Agents in shipcrew are designed to hand off to each other. Examples:
flake-hunterfinds flaky tests → emits a list of candidates → user feeds a candidate totest-failure-investigatorfor deep dive.release-notes-writerruns at release time → references PRs that were already reviewed bypr-reviewer.dep-updaterproduces a PR list → each PR is then reviewed bypr-reviewerbefore merge.
The handoffs are documented in each agent's prompt — Claude knows when to escalate.
PRs welcome. Read CONTRIBUTING.md. The bar:
- Solves a maintainer's problem (not a developer's — that's solo-skills territory).
- Is single-purpose and could not reasonably be split further.
- Has explicit "what NOT to do" rules.
- Emits a fixed output format, not vibes.
- Includes confidence calibration and abort conditions.
MIT — copy, modify, ship, sell. No attribution required, but appreciated.
- solo-skills — Claude Code skills for building/shipping. shipcrew is for maintaining.