CrewLoop

CrewLoop is a documentation-first framework of role-based AI skills. Each skill is a self-contained SKILL.md instruction set that agents load and follow, enforcing a structured workflow across discovery, architecture, design, implementation, review, and shipping.

Highlights

• Process-driven workflow: CrewLoop Hub, Architect, Designer, Engineer, Reviewer, Shipper, and thirteen supporting roles each own one phase and never invade another's territory.
• Mandatory specs: Every change, from a one-line fix to a full feature, gets a lightweight spec in specs/changes/ before implementation starts.
• Design before code: When there is UI, the Designer defines the aesthetic direction before the Engineer writes markup or styles.
• Docs by docs-writer: READMEs, module docs, and changelogs are owned by the docs-writer skill so the engineer can focus on code and tests.
• Quality gate: The Reviewer inspects every diff for spec compliance, security, performance, and AI artifacts before anything reaches the repository.
• Conventional Commits: The Shipper generates commit messages, branches, archives specs, and opens PRs following the Conventional Commits standard.

Quick Start

Install the CLI globally and load the full crew:

npm install -g @archznn/crewloop-cli
crewloop install

Install only the skills you need:

crewloop install --skill architect --skill engineer

Install to a custom directory or for another supported agent:

crewloop install --target /path/to/your/skills/dir
crewloop install --agent claude

Validate that all skills are well-formed:

python scripts/validate-skills.py

Each skill is automatically detected and activated according to the conversation context.

CLI Reference & Options

The crewloop CLI provides commands to manage skills and integrate them with your AI coding agents.

Commands

Global Flags for crewloop install

[!NOTE] DiamondBlock is opt-in. crewloop install never touches MCP configuration. To activate the optional DiamondBlock integration, install the official DiamondBlock CLI separately (for example npm i -g diamondblock) and run crewloop install --diamondblock, which delegates MCP registration to the official installer. Installing the DiamondBlock skill (Markdown instructions) is not the same as configuring or activating its MCP server. crewloop doctor reports DiamondBlock readiness as optional warnings, never errors.

Real-time Activity Dashboard

The dashboard provides a real-time WebSocket visualization of active skills, tool-use events, and execution logs.

By default, the dashboard binds to http://127.0.0.1:7890. You can change this port by setting the CREWLOOP_DASHBOARD_PORT environment variable.

Running the Dashboard

You can start the dashboard using the CLI:

crewloop dashboard

Alternatively, you can run it from the source:

cd servers/dashboard
npm install
npm run dev

Keyboard Shortcuts

• Cmd/Ctrl + K: Opens the command palette to search events, switch sessions, or manage active skills.

Supported Agents & Hooks

CrewLoop supports native shimming/hooking for the following AI agents:

• Kimi Code (kimi)
• Claude (claude)
• Codex (codex)
• AGY (agy)
• OpenCode (opencode)

During crewloop install, the installer modifies the configuration or custom scripts of the selected agent. This shims their execution, allowing tool execution events (such as read/write file, run command, etc.) to be forwarded to the local dashboard WebSocket.

Meet the Crew

CrewLoop ships 19 specialist skills. The core crew owns the main delivery loop; the supporting crew jumps in when the context demands it.

Core Crew

Supporting Crew

Skills in Action

Workflow (Direct Routing)

Skills hand off directly to the next skill via their ending menu; the user confirms each transition. The CrewLoop Hub mediates only as the entry point for new tasks and as the automatic router in AFK mode.

flowchart TD
    O["CrewLoop Hub\nEntry: Discovery & First Routing\n+ AFK Auto-Router"] --> A["Architect\nSpecs & Architecture"]
    A --> D["Designer\nUI/UX Direction"]
    A --> E["Engineer\nImplementation"]
    D --> E
    E --> R["Reviewer\nQuality Gate"]
    R -->|PASS| S["Shipper\nGit & PR"]
    R -->|FAIL| E
    S -->|new task| O

    PB["Project-Brainstorm\nDiscovery"] -.-> A
    LTM["Long-Term Manager\nMulti-Session Tracking"] <-.-> O
    DB["Diamondblock\nMemory & Distillation"] <-.-> O
    W["Docs-Writer\nDocumentation"] <-.-> O
    PM["Product-Manager\nPrioritization"] <-.-> O
    RS["Researcher\nTechnology Evaluation"] <-.-> O
    MN["Maintainer\nIncident & Debt"] -.-> A
    T["Tester\nQA Strategy"] <-.-> E

    A <-.-> SD["Schema-Designer\nAPI & DB Schemas"]
    D <-.-> FA["Frontend-Architect\nComponent Spec"]
    S <-.-> DO["DevOps-Specialist\nCI/CD & Docker"]
    R <-.-> SG["Security-Guard\nSecurity Review"]
    R <-.-> AA["Accessibility-Auditor\nAccessibility Review"]

Flow rules:

[!IMPORTANT] Core Routing Rule: Skills route directly to the next skill per the transition contract in references/conventions.md. The CrewLoop Hub mediates only at task entry and in AFK mode.

1. CrewLoop Hub is the entry point — it may invoke approved discovery/tracking helpers, then routes to Architect as the first mandatory delivery phase.
2. Architect is mandatory before implementation — Hub never routes directly to Designer or Engineer.
3. Architect is the design gatekeeper — once the spec is created, it recommends Designer (for UI) or Engineer (for code).
4. Designer acts before Engineer — when there is UI, the Designer creates the visual specification before the Engineer implements.
5. Engineer never does git, review, or docs — it implements code and tests, then its menu recommends the Reviewer.
6. Reviewer is the quality gate — no code reaches the repository without review. PASS recommends Shipper; FAIL recommends Engineer.
7. Shipper is the only skill that touches git — commit, branch, push, and PR. After shipping it offers a new task (CrewLoop Hub) or done.
8. Sub-skills assist core skills — supporting skills return to their invoker, except completed Project Brainstorm briefs and confirmed Maintainer bugs, which route to Architect.
9. Specs are archived — the specs/changes/ folder is moved to specs/archive/ on commit.
10. Bug-fixing Pipeline — Bug triaging is handled by the Maintainer, who recommends the Architect to create a lightweight specification (.spec.yaml + tasks.md), then the standard chain applies: Engineer → Reviewer → Shipper (commit/ship and archive the spec).
11. AFK mode is the exception — every non-Hub skill returns control to CrewLoop Hub, which loads the next skill without menus.

[!NOTE] Standard Developer Cycle Example: CrewLoop Hub (Discovery) -> Architect (Spec creation) -> Engineer (Build & Tests) -> Reviewer (Quality gate check) -> Shipper (Git commit & PR) -> done. Interactive transitions are confirmed through menus; Architect and Designer hand off automatically.

Repository Layout

crewloop/
├── skills/                # Role-based SKILL.md instructions
├── packages/cli/          # npm-published CLI installer
├── servers/dashboard/     # Real-time WebSocket dashboard
├── docs/                  # Docusaurus documentation site
├── references/            # Shared conventions and workflow reference
├── scripts/               # Validation and packaging helpers
└── specs/                 # Active, archived, and living specs

Adding a New Skill

1. Copy assets/templates/skill-template.md to skills/<skill-name>/SKILL.md.
2. Fill in the YAML frontmatter and role instructions.
3. Add the skill to the README tables if it is user-facing.
4. Run python scripts/validate-skills.py.
5. Open a PR; the Reviewer validates structure and the Shipper archives the spec.

Releasing

Versions are published automatically from main:

1. The Shipper bumps the version in package.json (and workspace manifests) following semver.
2. Merging to main triggers .github/workflows/release-tag.yml, which creates a vX.Y.Z tag.
3. .github/workflows/publish-npm.yml publishes @archznn/crewloop-skills to npm.

Manual releases are not required.

Contributing

Edit the files in skills/ and references/. Keep each SKILL.md concise and use reference files for shared detail. Run python scripts/validate-skills.py before opening a PR. For the full workflow, see references/workflow.md.

License

MIT