sumitaich1998/jarvisvr / skills/orchestration/task-decomposition
task-decomposition
VerifiedBreak a user's goal into a small DAG of sub-tasks for the JarvisVR specialist team, then emit it as an orchestration.plan. Use whenever a request needs more than one capability or agent, e.g. "what's on my desk and the weather in Tokyo, and start a 5-minute timer", multi-step errands, or any goal mixing perception, research, productivity, smart-home, navigation, media, comms, or stage work. Triggers: plan, decompose, break down, multi-step, "and also", "while you're at it", several things at once.
Task Decomposition
You are Jarvis, the L0 orchestrator. This skill turns one user goal into a
small, dependency-aware plan that the specialist roster can execute. You never do
domain work here — you only plan and publish the plan as orchestration.plan.
When to use
- Use when the goal needs two or more specialists, or one specialist across sequential steps with dependencies.
- Skip (single dispatch) when the goal is one obvious sub-task for one agent —
hand straight to
agent-routinginstead.
Steps
- Normalize the goal. Strip filler, keep intent + entities (cities, device
names, durations, people). If it is ambiguous and blocking, stop and run
clarify-intentfirst. - Enumerate atomic sub-tasks. One verb + one object each ("get Tokyo weather", "start 5-minute timer", "identify object on desk").
- Assign a role to each sub-task by matching it to the roster in
docs/ORCHESTRATION.md §3(do the actual matching withagent-routing). - Wire dependencies. Mark which sub-tasks are independent (run in parallel) vs. which consume another's output (sequence them). Keep the graph shallow.
- Allocate agent ids.
jarvisis L0; L1 agents area1, a2, …; sub-agents use dotted ids (a1.1). Keep depth ≤ 3. - Emit
orchestration.planonce, up front, then let the dispatcher driveorchestration.agent_statusper agent.
Output: orchestration.plan (protocol §9.2)
For "what's this on my desk, and the weather in Tokyo? Start a 5-minute timer."
{
"plan_id": "uuid-v4",
"goal": "identify the desk object, show Tokyo weather, start a 5-minute timer",
"agents": [
{ "agent_id": "jarvis", "role": "orchestrator", "name": "Jarvis", "parent": null, "level": 0 },
{ "agent_id": "a1", "role": "perception-agent", "name": "Perception", "parent": "jarvis",
"level": 1, "subtask": "identify the object on the desk", "skills": ["identify-object"] },
{ "agent_id": "a2", "role": "research-agent", "name": "Research", "parent": "jarvis",
"level": 1, "subtask": "get current weather for Tokyo", "skills": ["web-research"] },
{ "agent_id": "a3", "role": "productivity-agent", "name": "Productivity", "parent": "jarvis",
"level": 1, "subtask": "start a 5-minute timer", "skills": ["manage-timers"] },
{ "agent_id": "a4", "role": "stage-agent", "name": "Stage", "parent": "jarvis",
"level": 1, "subtask": "lay the results out comfortably", "skills": ["compose-workspace"] }
],
"edges": [
{ "from": "jarvis", "to": "a1" }, { "from": "jarvis", "to": "a2" },
{ "from": "jarvis", "to": "a3" }, { "from": "a1", "to": "a4" },
{ "from": "a2", "to": "a4" }, { "from": "a3", "to": "a4" }
]
}
a1–a3 are independent and run in parallel; a4 (stage) depends on all
three so it sequences last. Emit agent.thinking{stage:"planning"} before
the plan.
Dependency patterns
- Fan-out / parallel: independent reads (weather + identify + timer).
- Pipeline:
research-agent→summarize-source→stage-agent present-data. - Gather/synthesize: always route the final compositing to one
stage-agentnode so holograms don't collide (seeresult-synthesis).
Edge cases
- Single capability → don't emit a plan; dispatch one agent directly.
- Unknown / unsupported sub-task → drop it from the plan and note it for the spoken summary; never invent a role that isn't in the roster.
- Destructive or gated actions (unlock door, spend money) → keep them as their own node so consent can gate that step (perception/tool actions stay user-gated per v1.1).
- Too many sub-tasks (>6) → group related ones under one specialist subtask rather than spawning a wide, noisy graph.
- Cyclic dependency → you mis-modeled it; collapse the cycle into a single sequential node.
Hand-off
Pass each {agent_id, role, subtask} to the dispatcher; per-step routing detail
lives in agent-routing, and the final spoken/visual merge lives in
result-synthesis.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |