Skip to main content

Orchestrator Skill

The orchestrator skill enables automated multi-agent coordination for complex projects, managing subagent spawning and context-protected workflows.

Overview

Context Protection

Maintains 10K token budget by delegating all work

Wave Execution

Respects dependency order with wave-based spawning

Manifest-Based

Reads only summaries, never full file contents

HITL Friendly

Human stays in the loop for decisions

Installation

The orchestrator skill is included with CLEO’s skill system:
# Verify skill is available
cleo skill list | grep orchestrator

# Or activate via Claude Code
/ct-orchestrator

The Protocol

The orchestrator NEVER implements anything directly. It only plans, spawns, reads summaries, and coordinates.

The Mantra

Stay high-level. Delegate everything. Read only manifests. Spawn in order.
PhraseMeaningConstraint
Stay high-levelReason about what, never howORC-001
Delegate everythingZero code, zero file readingORC-002
Read only manifestsSummaries only, no full filesORC-003
Spawn in orderRespect dependency graphORC-004

Usage

/ct-orchestrator

"Orchestrate the authentication epic T001"

How It Works

Context Budget

The orchestrator operates within a strict context budget:
AllocationTokensPurpose
User conversation~5,000Dialogue and decisions
Task queries~2,000cleo show --brief
Manifest summaries~2,000Key findings
Planning/reasoning~1,000Wave computation
Total10,000Hard limit
Subagents get fresh 200K context each, enabling deep implementation work.

Manifest Protocol

Subagents write to MANIFEST.jsonl:
{
  "id": "task-T005-impl",
  "file": "research-outputs/2026-01-20_auth-impl.md",
  "title": "Authentication Implementation",
  "key_findings": [
    "JWT middleware implemented with refresh token support",
    "Password hashing uses bcrypt with cost factor 12",
    "Session binding integrated with CLEO"
  ],
  "needs_followup": ["T006"],
  "linked_tasks": ["T005"]
}
The orchestrator reads ONLY key_findings, never the full output file.

Wave Computation

Waves are computed from the dependency graph:
Wave 0: Tasks with no dependencies
Wave 1: Tasks depending only on Wave 0
Wave 2: Tasks depending only on Wave 0-1
...
1

Analyze Dependencies

cleo deps --json | analyze
2

Compute Waves

Group by dependency depth
3

Execute Wave

Spawn all tasks in current wave
4

Wait for Completion

Monitor subagent outputs
5

Advance

Move to next wave

Constraints Reference

IDConstraintViolation Consequence
ORC-001No implementationContext explosion
ORC-002Delegate via Task toolSubagent fresh context
ORC-003No file reads >100 linesBudget exceeded
ORC-004Spawn in orderWasted work
ORC-00510K context budgetProtocol broken

Best Practices

Break Down Epics

Create granular tasks before orchestrating

Define Dependencies

Clear dependency graph enables wave computation

Trust Subagents

Don’t micromanage - provide clear task descriptions

Read Summaries

Only read key_findings from manifest