| name | subagent-first |
| description | Route work through bounded subagents. Use when tasks require codebase investigation, code changes, tests, lint, build, or multi-step validation. |
Subagent-First
Overview
Main agent = orchestrator only.
It plans, routes, reviews concise results, and decides next step.
It does not read implementation files, implement, run tests, lint, or build.
It may read planning and design artifacts (plan files, ADR, docs/) directly for routing context.
Non-negotiable rules
- Do not delegate discovery + implementation + validation in one subagent.
- One subagent = one phase, one scope, one success signal.
Phases
Use separate subagents for these phases:
- discovery: find files, root cause, options. No edits.
- implementation: edit only explicitly named files.
- test: add or update focused tests only.
- validation: run requested tests/lint/build only.
- review: inspect final diff and risks. No edits unless asked.
Dispatch contract
Every subagent prompt must include:
- phase
- goal
- allowed files/dirs
- forbidden files/dirs
- edits allowed: yes/no
- success signal
- return format
Return format
Subagents return only:
status: success | failure | blocked
summary:
files_changed:
log_path:
next_action:
Verbose output goes to docs/tmp/.
Routing loop
After each subagent result, main agent must choose exactly one:
- dispatch next bounded subagent
- split scope
- ask user
- stop and report
Model
- haiku: discovery, validation, simple reads, running basic commands
- sonnet: implementation, root-cause analysis, review
