---
description: Orchestration - routes tasks to the right subagent automatically
alwaysApply: true
---

# Delegation

After reading memory-bank, check if the current task matches a subagent below. Delegate automatically; do not wait for user to name the agent.

## When to Delegate

<!-- Customize this table: replace rows with your project's agents and triggers -->

| Intent / Trigger | Subagent |
|---|---|
| Editing source code, implementation work in `{SOURCE_DIR}` | {CODE_AGENT} |
| "break this down", "scope", "plan feature", brainstorming, user stories, feature ideation, task decomposition | product-owner |
| After product-owner creates stories | priority-steward (schedule new tasks) |
| "what's next?", priorities, sprint, backlog, reprioritize | priority-steward |
| After completing a task or implementation | priority-steward (update Done, promote Next) |
| Code review, PR review, diff feedback | code-reviewer |
| Verify PRD compliance, trace FR-IDs, requirement gaps | prd-compliance |
| "update memory bank", after major implementation | memory-bank-updater |
| User gives process/workflow/style guidance | Handle directly: update relevant rules/skills/agents |

## Model Preferences

<!-- Adjust model preferences based on your available models -->

- **code-reviewer, {CODE_AGENT}**: Prefer code-focused model
- **product-owner, prd-compliance, memory-bank-updater**: Prefer reasoning model
- **priority-steward**: Default model is fine

## Rules

- Memory Bank rule takes priority over this rule when in conflict.
- All agents must follow `incremental-work.mdc`. Prefer small-to-medium incremental changes; large single-pass artifacts should be escalated to the user first.
- All agents may search the internet for additional context when project files are insufficient; prefer authoritative sources.
- All agents may use additional tooling, including command-line workflows and relevant CLIs, when needed to complete tasks.
- Do not delegate simple one-step tasks; handle them directly.
- When delegating, pass relevant context from the current conversation in the Task prompt. Subagents read only their domain-specific files (defined in each agent definition), not all memory-bank files.