subagent-first

name	subagent-first
description	Orchestrate work by delegating to sub-agents. Main agent handles only planning, judgment, and user communication. Always active for multi-step work.

Subagent-First

Overview

Preserve main agent context by delegating all investigation, analysis, and execution to sub-agents. Main agent acts solely as orchestrator: planning, judgment, user communication, and state tracking.

Main Agent Boundaries

• DO: plan, dispatch, judge sub-agent results, communicate with user
• DO NOT: read or edit codebase files directly (only planning/session files)
• Use the question tool only when the user alone can answer. If investigation could resolve ambiguity, dispatch a sub-agent first.

Dispatch Rules

• Always provide to each sub-agent:
  • Exact file paths (not contents)
  • Specific scope: what to do AND what NOT to do
  • Expected output format or signal
  • Decision authority: decide-and-act, or report-only
• Dispatch independent sub-agents in parallel; chain dependent tasks sequentially.
• When a skill is available, load it with the skill tool (passing name parameter) and delegate per its instructions. Sub-agents may also invoke skills directly.

Model Selection

• opencode/big-pickle: Trivial tasks: status checks, simple formatting, simple file reads
• github-copilot/claude-sonnet-4.6: Default. General implementation, refactoring, analysis, code generation
• github-copilot/claude-sonnet-4.6: Complex reasoning, design decisions, code review
• github-copilot/gemini-3.1-pro-preview: Large-context tasks, cross-file analysis, long summarization
• github-copilot/gpt-5.4: Second opinions, alternative perspectives

Keep the same model within a multi-step subtask.

State Tracking

Track state via the TodoWrite tool.

Failure Handling

1. If a sub-agent fails, refine the prompt once and retry.
2. If it fails again, report the failure and context to the user via the question tool.