parallel-dispatch

Parallel Dispatch

Safe parallel execution of multiple agents within the same domain. This skill defines the safety invariants — the constraints that, if maintained, guarantee conflict-free merges by construction.

Recursive by design. These invariants apply at every nesting depth. A @backend-engineer[auth] that spawns sub-subagents for [auth-login] and [auth-registration] must apply the same decomposition, ownership, and merge rules internally.

When to Invoke

• Before dispatching multiple instances of the same agent type within a primitive
• When any agent (at any depth) needs to parallelize its own workload across sub-subagents
• When re-planning after a sub-task failure

1. Decomposition

Break work into parallelizable, MECE sub-tasks. Each gets a scope card.

Protocol

1. Inventory — List all concrete deliverables (files, modules, endpoints, tests)
2. Cluster by feature — Group by vertical feature slice, not technical layer
3. MECE validate — No file in two clusters. No deliverable unassigned.
4. Dependency scan — Mark hard (blocks) vs soft (beneficial but not blocking) dependencies
5. Produce scope cards
6. Cap check — >5 instances of same agent type → prompt user

Critical anti-pattern: Layer-based decomposition (handlers/, services/, repos/) creates cross-cutting dependencies. Always use feature-based clustering (auth/, tasks/, notifications/).

Scope Card Format

### Scope Card: @<agent-type>[<scope>]
- **Deliverables**: <list of concrete outputs>
- **Write Scope**: <glob pattern for exclusive write access>
- **Shared Reads**: <glob patterns for read-only access>
- **Hard Dependencies**: <scope cards that must complete first>
- **Soft Dependencies**: <scope cards that are beneficial but not blocking>

Rules:

• Every card has at least one concrete deliverable
• Every card has a write scope (glob pattern)
• Write scopes between agents MUST be disjoint
• Scope names use kebab-case: [auth], [task-crud], [user-profile]
• Empty scope cards are invalid — merge with adjacent cluster