Task Orchestrator

SKILL.md

---
name: task-orchestrator
description: Plan and execute complex multi-step tasks using the agent team. Decomposes requests, routes to optimal agents, coordinates execution, compiles results, and stores knowledge in Obsidian. Use when a task requires multiple agents, research + action, or structured multi-step work.
metadata:
  openclaw:
    emoji: "🔄"
---

# Task Orchestrator

Coordinates complex multi-step tasks across the agent team.

## When to Use

✅ **USE this skill when:**
- A task needs multiple agents (research + code + images, etc.)
- Complex requests that need decomposition into steps
- Projects that generate knowledge worth storing
- Anything requiring a plan before execution

❌ **DON'T use for:**
- Simple one-shot questions (just answer directly)
- Single-agent tasks (just use that agent)
- Casual conversation

## Process

### Step 1 — Plan
Before executing anything, create a brief plan:

```
TASK: [Original request]
STEPS:
1. [Step] → Agent: [name] | Input: [what they need] | Output: [what they produce]
2. [Step] → Agent: [name] | Input: [what they need] | Output: [what they produce]
PARALLEL: [which steps can run simultaneously]
SEQUENTIAL: [which steps depend on prior outputs]
STORE: [what goes to Obsidian and where]
```

### Step 2 — Route
Use the agent routing matrix:

| Need | Agent |
|------|-------|
| Quick facts, current events | 📡 Sonar (perplexity/sonar-pro) |
| Deep research, multi-source | 🔍 Research Agent (claude-opus-4-6) |
| Code, build, debug | 💻 Codex (gpt-5.4) |
| Long autonomous tasks, batch work | 🟢 Nemotron (ollama/nemotron-3-super:cloud) |
| Image generation/editing | 🍌 Banana (nano-banana scripts) |
| Massive docs, video, audio analysis | 📚 Atlas (google/gemini-2.5-pro) |
| Location/business search | 📍 Google Places (skill script) |
| Orchestration, delivery | 🔪 Toji (me) |

### Step 3 — Execute
- Spawn agents using `sessions_spawn` for isolated work
- Use tools directly for quick lookups (web_search, exec)
- Run parallel where no dependencies exist
- Monitor long-running agents, don't poll in loops

### Step 4 — Compile
- Gather all outputs
- Synthesize into a coherent final deliverable
- Verify it answers the original request

### Step 5 — Store
Save valuable outputs to Obsidian:

```bash
# Research findings
obsidian-cli write "Research/[Topic]" --content "..."

# Decisions
obsidian-cli write "Decisions/$(date +%Y-%m-%d) [Title]" --content "..."

# References
obsidian-cli write "References/[Title]" --content "..."
```

Use frontmatter:
```yaml
---
tags: [relevant, tags]
created: YYYY-MM-DD
status: active
project: project-name
agents: [list, of, agents, used]
---
```

### Step 6 — Deliver
- Send result to David
- Update Mission Control if relevant (tasks, projects, activity)
- iMessage for anything urgent or requiring review

## Rules

- ALWAYS plan before executing multi-step work
- ALWAYS store research and decisions in Obsidian
- Prefer parallel execution where possible
- Don't use expensive models for cheap tasks (Sonar for quick facts, not Opus)
- Update Mission Control data when agents/projects/tasks change
- Notify David via iMessage for completed major work