Swarm Advisor

The swarm-orchestration skill contains the full technical primitives (Team, Teammate, Task, Inbox, Message). This skill adds the design layer — when to use a swarm, how to structure it, and how to avoid common traps.

---

Is a Swarm Worth It?

Ask first. Swarms have real overhead: spawn cost, coordination messages, inbox polling, cleanup.

Good fit:

• 3+ independent workstreams that can run in parallel
• Work that naturally divides: research/implement/test, feature-A/feature-B/feature-C
• Long-running tasks where parallelism saves real time
• Pipeline patterns where stage N feeds stage N+1 with clear handoffs

Not worth it:

• Single sequential task → just do it inline
• 2 tasks → use two Task() calls without a full team
• Tight interdependencies where agents constantly block each other
• Tasks requiring human judgment mid-way → use GSD interactive mode instead

---

Team Topology Patterns

Pattern 1: Parallel Workers (most common)

Leader
├── Worker A (area: auth)
├── Worker B (area: API)
└── Worker C (area: tests)

Use when: Independent features/modules, parallel research, parallel code review.

Pattern 2: Pipeline

Leader → Researcher → Synthesizer → Implementer → Verifier

Use when: Each stage's output feeds the next. Only parallelize within stages.

Pattern 3: Fan-out / Fan-in

Leader spawns N researchers in parallel
Leader waits for all
Leader synthesizes results

Use when: Exploring multiple approaches, broad codebase mapping, multi-source research.

Pattern 4: Self-organizing Queue

Leader creates task queue
Workers claim tasks
Workers report back
Leader assigns next task

Use when: Many small tasks of unknown count, load balancing needed.

---

Pre-flight: Task Decomposition

Before spawning anything, help the user decompose the work:

1. List all tasks — one sentence each
2. Mark dependencies — which tasks must finish before others start?
3. Group into waves — tasks with no blocking dependencies = Wave 1, etc.
4. Assign to agents — group related tasks per agent to minimize cross-agent state sharing

Present as a table:

Wave | Task | Depends on | Agent
-----|------|-----------|------
1    | Research auth patterns | — | researcher
1    | Research API patterns | — | researcher
2    | Implement auth | Wave 1 | implementer-A
2    | Implement API | Wave 1 | implementer-B
3    | Integration tests | Wave 2 | verifier

Ask: "Does this decomposition look right? Any tasks missing or misassigned?"

---

Status Tracking

While a swarm runs, periodically check:

# Check task statuses
cat ~/.claude/tasks/<team>/*/json 2>/dev/null | grep '"status"'

# Check for stuck agents (no progress in N messages)
cat ~/.claude/teams/<team>/inboxes/<agent>.json

Signs of a stuck swarm:

• Tasks stuck in_progress for many iterations
• Agents sending idle_notification repeatedly
• Inbox messages going unread

Recovery: Send a targeted message to the stuck agent with clarification, or reassign the task.

---

Shutdown Checklist

Before declaring done:

• All tasks in completed status
• Leader has received results from all teammates
• No pending shutdown_request messages unacknowledged
• Run cleanup: TeamDelete to remove team config

---

Integration with Other Frameworks

Use swarm when...	Combined with
GSD execute-phase has 3+ independent plans	GSD spawns agents via swarm pattern
Gepetto section files are independent	Swarm implements sections in parallel
Ralph-loop hits a complex multi-part task	Pre-split into swarm, then loop on each part