swarm-coding-skill

# Swarm Coding Skill

Fully autonomous multi-agent software development. Given a plain-English prompt, the swarm designs, implements, tests, and delivers a complete project end-to-end.

**Core capability:** Code generation via OpenRouter's qwen3-coder model. The orchestrator drives a Planner to create a manifest, then executes specialized worker roles (BackendDev, FrontendDev, QA, DevOps, etc.) in dependency order. All code is written to files; no interactive sessions.

**Important:** This skill **generates code** for review and deployment by the user. It does not make business decisions or operate autonomously in production. The user remains responsible for security, compliance, and operational decisions.

## How It Works

1. **Orchestrator** (`Planner` role) analyzes your prompt, decides tech stack and architecture, and creates a `swarm.yaml` manifest with tasks and dependencies.
2. **Worker agents** (`BackendDev`, `FrontendDev`, `QA`, `DevOps`) are spawned as sub-sessions. Each has a clear persona and works on its assigned files in a shared workspace.
3. **Coordination**: The orchestrator tracks task completion and dependencies. When a task finishes, it marks it done and starts any unblocked downstream tasks.
4. **Conflict avoidance**: Files are partitioned by role (Backend owns `server/`, Frontend owns `client/`, etc.). If two roles need the same file, the manifest assigns an owner.
5. **Quality gates**: QA must pass tests before integration; DevOps ensures containerization; no merge without green tests.
6. **Deliverable**: You get a complete project directory with README, tests, Dockerfile, and optionally a GitHub repo or zip.

## Usage

```bash
# In your main OpenClaw session, invoke:
/trigger swarm-code "Build a dashboard that shows Moltbook stats and ClawCredit status"
```

The skill will:
- Spawn the orchestrator in an isolated session
- Orchestrator spawns workers sequentially or in parallel (based on dependencies)
- Output a final summary and path to the...

# Swarm Coding Skill

Fully autonomous multi-agent software development. Give a plain-English prompt describing an app, and the swarm designs, implements, tests, and delivers a complete project end-to-end.

## Requirements

- Node.js v18+
- OpenRouter API key with access to `qwen/qwen3-coder`
- Workspace with `.env` containing `OPENROUTER_API_KEY`

## Usage

```bash
# From the workspace root, run:
cd clawhub-swarm-coding-skill
node orchestrator.js "Build a dashboard that shows Moltbook stats and ClawCredit status"
```

The orchestrator will:
1. Create a project workspace under `swarm-projects/<timestamp>/`
2. Spawn a Planner (uses qwen-coder) to generate a `swarm.yaml` manifest
3. Execute specialized workers: BackendDev, FrontendDev, QA, DevOps
4. Assemble final project in the same directory
5. Print the final location on success

You can then `cd` into that directory and run `npm install` and `npm start`.

## How It Works

- **Exclusive model**: All agents use `qwen/qwen3-coder` (the correct OpenRouter model ID is `qwen/qwen3-coder`, *not* `openrouter/qwen/qwen3-coder`).
- **Manifest**: A YAML-like JSON describing roles, outputs, and dependencies
- **Workers**: Each runs sequentially based on dependencies; each writes files to a shared `files/` workspace
- **Retry**: Failed tasks are retried once with stricter instructions
- **Quality**: QA must create tests; DevOps adds Docker and CI; all code is production-style

### Model Access

Ensure your OpenRouter API key has `qwen/qwen3-coder` enabled:
1. Go to https://openrouter.ai/keys
2. Edit your key
3. Under "Model Access", check `qwen/qwen3-coder`
4. Save

If you get `"not a valid model ID"` errors, it usually means the model ID format is wrong or your key lacks access.

## Output Structure

```
swarm-projects/
└── swarm-2026-02-14T21-30-00/
    ├── swarm.yaml              # manifest
    ├── tasks.json              # execution log
    ├── DECISIONS.md            # architectural decisions with rationale
    ├── .learn...