autopilot

Autopilot — Fully Autonomous Execution Mode

Autopilot mode handles a development task end-to-end without user intervention.

Workflow

Step 1: Gather Requirements

Use AskUserQuestion to understand the task:

{
  "questions": [
    {
      "question": "What task should autopilot handle?",
      "header": "Task",
      "options": [
        {"label": "Implement feature", "description": "Build a new feature from scratch"},
        {"label": "Fix bugs", "description": "Diagnose and fix one or more bugs"},
        {"label": "Refactor code", "description": "Restructure existing code without changing behavior"},
        {"label": "Full cycle", "description": "Plan → Implement → Test → Review → Document"}
      ],
      "multiSelect": false
    },
    {
      "question": "What quality level is required?",
      "header": "Quality",
      "options": [
        {"label": "Production (Recommended)", "description": "Full tests, code review, documentation. Uses Sonnet/Opus."},
        {"label": "Prototype", "description": "Working code, minimal tests. Uses mostly Haiku/Sonnet."},
        {"label": "Maximum", "description": "Production + security review + critic review. Uses Opus heavily."}
      ],
      "multiSelect": false
    }
  ]
}

Step 2: Plan

Dispatch the planner agent (Opus) to:

1. Analyze the codebase context
2. Decompose the task into subtasks
3. Assign tiers and agents to each subtask
4. Identify parallel execution opportunities
5. Define quality gates

Step 3: Execute

Based on the plan, dispatch agents in the optimal order:

Parallel phase (independent tasks):

• Dispatch multiple agents simultaneously using parallel Task calls
• Explorer (Haiku) for codebase scanning
• Researcher (Sonnet) for technical investigation if needed

Sequential phase (dependent tasks):

• Executor (Sonnet) for implementation
• Build-fixer (Sonnet) if build errors occur
• TDD-guide (Haiku) for test writing

Step 4: Quality Gates

After implementation, run quality checks:

• Code Review: Dispatch code-reviewer (Sonnet) on all changed files
• Tests: Run test suite via Bash
• Build: Verify clean build
• Security (if Maximum quality): Dispatch security-reviewer (Opus)