orchestration

Orchestration Skill — Task Decomposition + Parallel Coordination

Purpose

This skill decomposes large tasks into independent units and coordinates parallel execution. It is a PLANNING + COORDINATION tool — it does NOT spawn agents directly. Agents are spawned via Claude Code's built-in /batch command (git worktree-based parallel).

This skill answers: "How do I split this work? Which agents do what? How do I make sure they don't conflict? How do I merge their outputs?"

---

---

How to Run

run_orchestration(task, {"phase": "analyze"})   → Decompose task, score parallelism
run_orchestration(task, {"phase": "plan"})     → Present plan (after analyze)
run_orchestration(task, {"phase": "execute"})  → Build unit configs → use /batch to spawn
run_orchestration(task, {"phase": "update", "unit_id": "...", "result": {...}})  → Record unit result
run_orchestration(task, {"phase": "merge"})    → Validate + merge results
run_orchestration(task, {"phase": "report"})    → Final summary

Execute Phase — neutron_spawn_agent

The execute phase returns unit configs. Call neutron_spawn_agent for each unit via MCP:

1. run_orchestration(..., phase="execute") → unit configs in output
2. For each unit: call neutron_spawn_agent(agent_id=unit_id, prompt=unit_prompt, ...)
   → agents run in parallel via claude-agent-sdk
3. Wait for results (or use background=True for non-blocking)
4. run_orchestration(..., phase="merge", results={...}) → validate + merge

Each spawned agent uses Claude Code's tools (Read/Edit/Bash/Glob/Grep) with its own context window. Agents can run in the same project or different subdirectories (use cwd param to isolate).

# After run_orchestration(..., phase="execute"):
configs = result["agent_configs"]  # list of agent configs

# Call Agent() for EACH config concurrently:
for cfg in configs:
    Agent(
        prompt=cfg["prompt"],
        agent=cfg["agent"],        # "Plan" | "Explore" | "general-purpose"
        background=cfg["background"],  # True = concurrent, False = blocking
        maxTurns=cfg["max_turns"],
        isolation=cfg["isolation"],   # project root path → git worktree isolation
        skills=cfg["skills"],         # ["spec", "context"] preloaded
    )

# When ALL agents finish, call:
run_orchestration(task, {"phase": "merge"})