// Decomposes broad tasks into MECE, parallelizable sub-tasks with explicit scope cards. Core skill for intra-domain parallel dispatch. Produces scope cards consumed by parallel-dispatch-dag, parallel-dispatch-ownership, and parallel-dispatch-merge skills.
| name | parallel-dispatch-decomposition |
| description | Decomposes broad tasks into MECE, parallelizable sub-tasks with explicit scope cards. Core skill for intra-domain parallel dispatch. Produces scope cards consumed by parallel-dispatch-dag, parallel-dispatch-ownership, and parallel-dispatch-merge skills. |
Break a broad task into parallelizable, MECE sub-tasks. Each sub-task gets a scope card that defines its boundaries, deliverables, and dependencies.
| Phase | Decomposition Axis | Example Scopes |
|---|---|---|
| SCOUT | Investigation area (feature, subsystem, technology) | [auth], [tasks], [notifications] |
| DESIGN | Architectural concern (contracts, data model, boundaries) | [api-contracts], [data-model], [component-boundaries] |
| PRE-MORTEM | Risk domain (failure modes, blast radius, recovery) | [failure-modes], [blast-radius], [recovery-paths], [threat-model] |
| BUILD | Feature slice (vertical feature boundaries) | [auth], [tasks], [lists] |
| TEST | Test suite domain (E2E suite, contract suite) | [auth-e2e], [task-e2e], [api-contract] |
| REVIEW | Review dimension or feature scope | [auth-review], [security-dim], [test-coverage] |
| REMEDIATE | Mirrors BUILD decomposition (fix what you built) | Same scopes as BUILD findings |
| OPTIMIZE | Performance domain (CPU, memory, I/O, queries) | [query-opt], [cache], [concurrency] |
| REFACTOR | Module boundary or code smell cluster | [user-module], [task-module] |
List all concrete deliverables for this primitive. Be specific: files, modules, endpoints, tests.
Group deliverables by natural boundaries. Prefer feature-based clustering over technical-layer clustering.
Correct (feature-based):
Cluster: auth → handler, service, repository, middleware, tests
Cluster: tasks → handler, service, repository, tests
Wrong (layer-based):
Cluster: handlers → auth handler, task handler, list handler
Cluster: services → auth service, task service, list service
Feature-based clustering produces naturally independent sub-tasks. Layer-based creates cross-cutting dependencies.
If either check fails → re-cluster.
For each cluster, identify:
Mark dependencies as:
hard → cannot start without this input (blocks parallel execution)soft → benefits from this input but can proceed with assumptions (parallelizable)Produce one scope card per sub-task:
### Scope Card: @<agent-type>[<scope>]
- **Deliverables**: <list of concrete outputs>
- **File Allowlist**: <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>
- **Blocked By**: <list of scope card names>
- **Blocks**: <list of downstream scope card names>
Count sub-tasks per agent type. If >5 instances of the same agent type:
[auth], [task-crud], [user-profile]| Anti-Pattern | Why It Fails | Fix |
|---|---|---|
| Layer-based decomposition | Creates cross-cutting deps, every sub-task touches every feature | Use feature-based clustering |
| Shared write access | Two agents editing same file → merge conflicts guaranteed | One writer per file, extract shared code to contract layer |
| Scope too small (<1 file) | Overhead exceeds benefit, context fragmentation | Merge into adjacent cluster |
| Scope too broad (entire domain) | No parallelism gained, single bottleneck | Break into feature slices |
| Missing MECE validation | Gaps → missed deliverables, overlaps → conflicts | Always run Step 3 |
| Decomposing without design output | No contracts → agents make incompatible assumptions | DESIGN before BUILD decomposition |
Read-only agents (scout, qa-analyst, security-engineer, ux-reviewer, incident-responder) use MECE scoping for coverage guarantees, not conflict prevention:
Some files inherently aggregate contributions from multiple features (routers, registries, main entry points, config files). Handle with a dedicated integration sub-task:
### Scope Card: @<agent-type>[integration]
- **Deliverables**: Wire all feature modules into application entry point
- **File Allowlist**: <aggregation files only>
- **Hard Dependencies**: All parallel sub-tasks for this agent type
- **Blocks**: REVIEW phase
Integration sub-tasks always run after all parallel sub-tasks complete. They merge last.
Session bootstrap + workflows for Pathfinder semantic navigation tools. Covers: discovery protocol, tool chaining patterns (explore, impact, audit, debug), search optimization, LSP degraded mode, and error recovery.
Playwright browser automation via MCP. Covers E2E testing, UI review, web scraping, screenshot capture, and general browser interaction. MCP-first — CLI is fallback only.
Safe command execution: input sanitization, timeout handling, output capture, error propagation. For spawning processes, shell commands, system calls.
Git conventions: conventional commits, branch naming, PR hygiene, release tagging.
Structured incident workflow: severity classification, triage, diagnosis, mitigation, postmortem, and prevention. Template-driven with blameless review.
Constructs, validates, and traverses a Directed Acyclic Graph (DAG) from scope cards for safe level-based parallel dispatch. Determines execution order via topological sort. Detects cycles and invalid dependencies.