| name | subagent-dispatch |
| description | Dispatch subagents for parallel work. Triggers on: 'dispatch subagents', 'use subagents', 'parallel execution'. |
Subagent Dispatch
Execute an implementation plan by dispatching independent subagents for each task, with two-stage review for quality.
⛔ REQUIRES a user-approved plan before starting. Never dispatch subagents without explicit consent.
Prerequisites
- An approved implementation plan exists (created via
writing-plans or equivalent)
- Plan has been reviewed and approved by the stakeholder
- Each task in the plan is independent enough to be executed by a single subagent
Step 1: Read the Plan
Load the approved plan and identify all tasks. Each task should have:
- Clear scope and deliverables
- File paths to create or modify
- Acceptance criteria
- Dependencies on other tasks (if any)
Step 2: Dispatch Subagents
For each independent task:
- Create directory structure if needed (
mkdir -p)
- Pre-create empty files if the subagent needs to create new files (custom agents lack
create tool)
- Dispatch one subagent per task with full context:
- What to implement
- Which files to modify
- Acceptance criteria from the plan
- Any constraints or conventions
Rules:
- One task per subagent — never bundle multiple tasks
- Include full context — subagents have no memory of previous tasks
- Specify unique names for classes/functions to avoid conflicts
- Run independent tasks in parallel when possible
Step 3: Two-Stage Review
After each subagent completes:
Stage 1: Spec Compliance
- Does the output match the plan's acceptance criteria?
- Are all required files created/modified?
- Does it integrate correctly with other completed tasks?
Stage 2: Code Quality
- Run linter: check for style violations
- Run type checker: verify type annotations
- Run tests: ensure nothing is broken
- Review for readability and maintainability
Step 4: Commit
After all tasks pass review:
git add -A
git commit -m "feat: [description from plan]
Refs #<issue-number>"
Rules
- Never dispatch without an approved plan — this is non-negotiable
- One task per subagent — keeps scope clear and failures isolated
- Verify output independently — don't trust agent claims, check the diff
- Pre-create files for custom agents — they can only edit existing files
- Full context in every prompt — subagents are stateless