drupal-delegation

Drupal Delegation

Parallel execution of implementation plans using the plugin's specialized agents.

When to Use

• You have a plan with 2+ independent implementation tasks
• Multi-file or multi-module changes
• After drupal-brainstorming or writing-plans produced a plan
• The plan has tasks that can be worked on without shared state

The Delegation Pattern

Step 1: Check Preconditions

Before dispatching agents:

1. Plan exists — Do NOT delegate without a plan. If no plan exists, use drupal-brainstorming first
2. Structural index is fresh — Run /drupal-status to check. Agents need accurate structural data
3. Tasks are independent — Each agent task must be completable without waiting for another agent's output

Step 2: Dispatch Specialized Agents

Use the right agent for each task type:

Task Type	Agent	Model	When
New code, services, entities, hooks	@drupal-builder	opus	Implementation work
Architecture review, security audit	@drupal-reviewer	sonnet	After implementation
Smoke tests, drush eval, config checks	@drupal-verifier	sonnet	After implementation
Semantic docs, tech specs	@semantic-architect	sonnet	After structural changes

Dispatch in parallel when tasks are independent:

Agent({ subagent_type: "drupal-workflow:drupal-builder", description: "Implement X", prompt: "..." })
Agent({ subagent_type: "drupal-workflow:drupal-builder", description: "Implement Y", prompt: "..." })

Dispatch sequentially when one depends on another:

1. @drupal-builder implements the feature
2. @drupal-reviewer reviews the implementation
3. @drupal-verifier verifies it works

Step 3: Write Agent Prompts

Each agent prompt MUST be self-contained. The agent has NO context from this conversation.

Include in every prompt:

• What to implement (exact requirements)
• Where (file paths from the plan)
• How (specific patterns: DI, #[Hook] attributes, accessCheck)
• Verification (how to confirm it works)

Bad prompt: "Implement the feature from the plan" Good prompt: "Create a service at code/my_module/src/Service/Calculator.php that implements CalculatorInterface with method add(int, int): int. Register in my_module.services.yml with interface type-hint. Verify with: ddev drush eval '...'"

Step 4: Track Progress

Create a task for each delegated unit:

TaskCreate({ description: "Implement Calculator service" })
TaskCreate({ description: "Implement admin form" })
TaskCreate({ description: "Verify all services" })

Update tasks as agents complete. This creates a visible progress trail.

Step 5: Verify After All Agents Complete

After all agents finish:

1. Review outputs — read what each agent produced
2. Dispatch @drupal-verifier — verify the combined result works
3. Check blast radius — run discover deps:FEATURE to confirm no unintended side effects

Dispatch Rules

• Use run_in_background: true for agents that don't block subsequent work
• Never dispatch more than 4 agents simultaneously — diminishing returns on context
• Include structural context in prompts when agents need to understand dependencies
• Do NOT delegate understanding — you synthesize results, agents execute specific tasks

Anti-Patterns

• Delegating without a plan (random agent dispatches)
• Using general-purpose agents for Drupal work (use specialized agents)
• Writing vague prompts ("based on the plan, implement it")
• Skipping verification after agent completion
• Dispatching dependent tasks in parallel

Relationship to Other Skills

drupal-brainstorming → writing-plans → drupal-delegation → verification-before-completion
        ↑                                      ↓
        └─── If delegation reveals gaps ───────┘