mobilize

name	mobilize
description	Multi-agent coordination protocol for tasks spanning 3+ distinct domains. Decomposes goal into sub-tasks, spawns specialized agents in isolated worktrees, aggregates results.
command	/mobilize
allowed-tools	["Bash","Read","Write","Glob","Grep"]
user-invocable	true

/mobilize — Multi-Agent Coordination Protocol

Trigger Conditions

• User invokes /mobilize
• User says "coordinate" + describes 3+ distinct domains
• Task touches 3+ unrelated domains simultaneously

Protocol

Step 1 — Decompose (Orchestrator, Plan Mode)

1. Enter plan mode immediately.
2. Break the goal into sub-tasks, one per domain/concern.
3. Write decomposition to tasks/todo.md under Queued.
4. Write agent assignments to tasks/status.md (format below).
5. Maximum 5 parallel agents.

Step 2 — Spawn Agents

For each sub-task:

• Spawn a Task agent with the appropriate agent type (see CLAUDE.md Delegation table).
• Pass only the minimum context needed for that sub-task.
• Every agent gets an isolated worktree (isolation: worktree).
• Each agent writes its result to tasks/results/{agent_name}-{YYYY-MM-DD}.md.
• Agents NEVER write to tasks/todo.md.

Step 3 — Monitor

• Poll tasks/results/ until all expected result files appear.
• Update tasks/status.md as agents complete.
• If an agent fails or times out, note in tasks/status.md and re-spawn once.

Step 4 — Aggregate

• Read all result files from tasks/results/.
• Synthesize into a final summary for the user.
• Move completed tasks in tasks/todo.md from Queued/In Progress → Done.
• Clean up tasks/status.md.

---

tasks/status.md Format

# Mobilize Status — {goal summary}
Started: {timestamp}

| Agent | Sub-task | Status | Result file |
|-------|----------|--------|-------------|
| python-specialist | Refactor X | done | tasks/results/python-specialist-2026-03-11.md |
| code-reviewer | Review PR | running | — |

---

Constraints

• Max 5 parallel agents.
• Each agent: isolated worktree at /tmp/dqiii8-wt/{agent_id}.
• Orchestrator does NOT share its context window with subagents.
• Destructive actions always require user confirmation before spawning agents.
• If decomposition is unclear, ask the user ONE clarifying question before proceeding.