implement-swarm

Implement Swarm Command

Execute implementation from conversation context using parallel workers. All workers complete → done.

Note: Loop and swarm are interchangeable - swarm is just faster when tasks can run in parallel. Both enforce exit criteria and sync state.

Source: Conversation context + argument input + mentioned files.

Uses Claude Code's built-in Task Management System for dependency tracking and visual progress (ctrl+t).

Arguments

• <task description> (required): What to implement
• --workers N (optional): Max concurrent workers (default: 3)
• --model MODEL (optional): Model for workers: haiku, sonnet, opus (default: sonnet)

Instructions

Step 1: Create Task Graph Immediately

ONLY use what's already available:

• The user's argument input
• Conversation history (already in context)

DO NOT:

• Read files unless the user explicitly asks you to
• Grep or explore the codebase
• Use Glob to find files

Create tasks immediately from context. Include file paths in task descriptions so workers can read them during execution.

For each work item, create a task:

TaskCreate({
  "subject": "Fix auth token validation",
  "description": "Full implementation details from context",
  "activeForm": "Fixing auth token validation"
})

Set dependencies with addBlockedBy — identify which tasks depend on others completing first:

// Task 2 needs the type fix from task 1
TaskUpdate({
  "taskId": "2",
  "addBlockedBy": ["1"]
})

A task with non-empty blockedBy shows as blocked in ctrl+t. When a blocking task is marked completed, it's automatically removed from the blocked list. A task becomes ready when its blockedBy list is empty.

Step 2: Spawn Workers

Worker limit N = --workers value or 3 if not specified. This is a queue — spawn up to N, then wait for completions before spawning more.

Mark each task in_progress before spawning its worker. Spawn up to N background workers in a SINGLE message (all Task calls in one response):