What is this?
CDO (Complexity-Driven Orchestration) is a graph-based execution system where agents read/write to disk in structured workflows. Problems are classified by complexity, then routed to appropriate patterns (Quick/Base/Deep/Epic).
Why does it exist?
Traditional linear thinking fails for interconnected problems. CDO prevents groupthink by ensuring agents never see each other's work during execution - only the final synthesis step reads all artifacts together.
When to use it:
Design work - Architecture decisions, system design
Complex debugging - Root cause analysis with multiple hypotheses
Strategic analysis - Multi-perspective evaluation of decisions
Assumption exploration - Socratic questioning of beliefs
Research synthesis - Integrating findings from multiple sources
When NOT to use it:
Simple questions with known answers → Use Quick workflow or just answer
Single perspective sufficient → Direct execution
No complexity/interdependencies → Standard task execution
CLI Commands
Primary Commands
# Auto-classify and route to appropriate workflow
lev cdo "{query or problem statement}"# Force specific complexity level
lev cdo --complexity=quick "{query}"
lev cdo --complexity=base "{query}"
lev cdo --complexity=deep "{query}"
lev cdo --complexity=epic --bd-epic=clawd-xxx "{query}"# Classification only (no execution)
lev cdo classify "{input}"# Debug mode (RCA workflow)
lev cdo debug "{error description}"
Examples
# Example 1: Auto-classify (likely: base)
lev cdo "Should we use GraphQL or REST for the API?"# Example 2: Force deep complexity
lev cdo --complexity=deep "Explore assumptions behind microservices"# Example 3: Debug workflow
lev cdo debug
lev cdo --complexity=epic --bd-epic=clawd-042
"Gateway crashes on Telegram messages"
# Example 4: Epic with BD tracking
"Evaluate voice assistant architecture"
Workflows
Overview: Four Complexity Levels
Level
Agents
Turns
Pattern
Use When
Quick
1-2
1
Sequential
Simple question, single perspective
Base
2-3
2
Fan-out/merge
Need 2+ perspectives, then synthesis
Deep
3-5
3-5
Multi-turn chains
Root cause analysis, assumption drilling
Epic
5+
5-10
BD-tracked phases
Strategic analysis, multi-session work
Routing logic:
Confidence ≥0.90 → Quick (direct execution)
Confidence ≥0.80 → Base (fan-out perspectives → synthesis)
Confidence ≥0.60 → Deep (multi-turn chains with convergence)
Confidence <0.60 → Epic (BD-tracked, human-in-loop)
Workflow 1: Classification & Routing
Use case: First step of every CDO invocation
Handled by:skill://lev-cdo/router
Process:
Parse input for intent (question, command, idea, research)
Assess confidence (0.0-1.0)
Determine complexity (quick, base, deep, epic)
Validate DoR (Definition of Ready)
Route to appropriate workflow
DoR Gates:
Sanity check (problem/solution alignment)
Prior art check (search for existing implementations)
Role definition - Clarifies operating scope and prevents ambiguous execution.
Context enrichment - Captures required inputs before actions.
Output structuring - Standardizes deliverables for consistent reuse.
Step-by-step workflow - Reduces errors by making execution order explicit.
Edge-case handling - Documents safe fallbacks when assumptions fail.
Technique Notes
These techniques improve reliability by making intent, inputs, outputs, and fallback paths explicit. Keep this section concise and additive so existing domain guidance remains primary.
Prompt Architect Overlay
Role Definition
You are the prompt-architect-enhanced specialist for lev-cdo, responsible for deterministic execution of this skill's guidance while preserving existing workflow and constraints.
Input Contract
Required: clear user intent and relevant context for this skill.
Preferred: repository/project constraints, existing artifacts, and success criteria.
If context is missing, ask focused questions before proceeding.
Output Contract
Provide structured, actionable outputs aligned to this skill's existing format.
Include assumptions and next steps when appropriate.
Preserve compatibility with existing sections and related skills.
Edge Cases & Fallbacks
If prerequisites are missing, provide a minimal safe path and request missing inputs.
If scope is ambiguous, narrow to the highest-confidence sub-task.
If a requested action conflicts with existing constraints, explain and offer compliant alternatives.
