lev-cdo-workflows

name	lev-cdo-workflows
description	[WHAT] Four complexity levels of CDO workflows: Quick, Base, Deep, Epic [HOW] Graph-based agent dispatch with disk-based artifact passing (no groupthink) [WHEN] After classification determines complexity level [WHY] Different problems need different depth - prevents over/under-engineering Triggers: Called by lev-cdo router based on complexity classification
skill_type	playbook
category	process-thinking
lifecycle_integration	{"stage":"crystallizing","input_artifact":"classification metadata + user query","output_artifact":"FINAL-synthesis.md (workflow output)"}
related_skills	["skill://lev-cdo/router","skill://lev-cdo/skill-discovery","skill://bd"]
plankton	false

CDO Workflows - Complexity-Based Execution Patterns

Lev Concept

What is this? CDO workflows are graph-based execution patterns where agents read/write to disk in sequence. Each complexity level (Quick/Base/Deep/Epic) defines agent count, turn structure, and convergence criteria.

Why does it exist? Graph structure + disk I/O prevents groupthink. Agents never see each other's work during execution - only the FINAL synthesis step reads all artifacts together.

When to use it:

• Quick: Simple questions, 1-2 agents, single perspective
• Base: Multi-perspective analysis, 2-3 agents, fan-out/merge
• Deep: Root cause analysis, 3-5 agents, multi-turn chains
• Epic: Strategic decisions, 5+ agents, BD-tracked multi-session

---

Core Principle

Graph-based layouts + Agentic execution = CDO

NOT: streaming, parallel, adaptive "properties" (fever dream) YES: YAML describes structure → Agents read/write disk → No groupthink

The Pattern That Works

./tmp/<workflow>-<timestamp>/
├── 00-input.md           # Original prompt/belief
├── 01-step-one.md        # Agent A output (reads: 00)
├── 02-step-two.md        # Agent B output (reads: 01)
├── 03a-turn-1.md         # Multi-turn: Agent C turn 1
├── 03b-turn-2.md         # Multi-turn: Agent C turn 2
├── 03c-turn-3.md         # Multi-turn: Agent C turn 3
├── 03-step-three.md      # Agent C synthesis
└── FINAL-synthesis.md    # ONLY this reads ALL files

Critical rules:

1. Each agent reads from DISK (previous step file)
2. Each agent writes to DISK (its output file)
3. Agents NEVER see each other's work during execution
4. Orchestrator ONLY launches agents—never synthesizes itself
5. Only FINAL step reads all artifacts together

---

CLI Commands

Primary Commands

# Execute workflow by complexity (called by router)
lev cdo execute --complexity=quick --input="00-input.md"
lev cdo execute --complexity=base --input="00-input.md"
lev cdo execute --complexity=deep --input="00-input.md"
lev cdo execute --complexity=epic --bd-epic=clawd-xxx

# Resume interrupted workflow
lev cdo resume --workflow-dir=tmp/analysis-20260128-143022

Execution Protocol

# 1. Create artifact directory
mkdir -p tmp/<workflow>-$(date +%Y%m%d-%H%M%S)

# 2. Write input file
echo "User's question or belief" > tmp/<workflow>-*/00-input.md

# 3. Launch agents per turn (use Task tool)
# Each agent reads previous step, writes its step

# 4. Final synthesis reads all

---

Workflows

Level 1: Quick (1-2 agents, 1 turn)

Use case: Simple question, single perspective, known answer pattern

Complexity factors: 0-2

• Single question
• Known answer pattern
• No dependencies

Pattern: Sequential execution, direct answer

Example:

Prompt: "Summarize this document"

workflow: quick-summary
agents:
  - step: 01
    skill: summarize
    input: "00-input.md"
    output: "FINAL-summary.md"

Execution:

Agent: Read 00-input.md → Write FINAL-summary.md
Done.

Convergence: N/A (single agent, single output)

---

Level 2: Base (2-3 agents, 2 turns)

Use case: Need 2+ perspectives, then synthesis

Complexity factors: 3-5

• Multiple perspectives needed
• 2-3 related domains
• Some prior art exists

Pattern: Fan-out/merge (parallel perspectives → synthesis)

Example:

Prompt: "Analyze pros and cons of X"

workflow: pros-cons-analysis
turns:
  - turn: 1
    parallel: true
    agents:
      - step: 01a
        skill: argue-for
        input: "00-input.md"
        output: "01a-pros.md"
      - step: 01b
        skill: argue-against
        input: "00-input.md"
        output: "01b-cons.md"

  - turn: 2
    agents:
      - step: 02
        skill: synthesize
        input: ["01a-pros.md", "01b-cons.md"]
        output: "FINAL-analysis.md"

Execution:

Turn 1 (parallel):
  Agent A: Read 00-input.md → Write 01a-pros.md
  Agent B: Read 00-input.md → Write 01b-cons.md
  ⏱ Sync

Turn 2:
  Agent C: Read 01a + 01b → Write FINAL-analysis.md
Done.

Convergence: Synthesis agent reads all Turn 1 outputs, no iteration needed

---

Level 3: Deep (3-5 agents, 3-5 turns)

Use case: Root cause analysis, debate, drilling into assumptions

Complexity factors: 6-9

• Root cause unknown
• Multiple hypotheses
• Cross-domain analysis
• 5+ agents needed

Pattern: Multi-turn chains with optional resonance loops

Example: Axiom Exploration Pattern

Prompt: "Explore foundational assumptions behind belief X"

workflow: axiom-exploration
skill_ref: workshop/poc/skills/domains/axioms/
turns:
  - turn: 1
    agents:
      - step: 01
        skill: paraphrase-engineer
        input: "00-input.md"
        output: "01-paraphrase.md"

  - turn: 2
    agents:
      - step: 02
        skill: steelman-enhance
        input: "01-paraphrase.md"
        output: "02-steelman.md"

  - turn: 3  # MULTI-TURN AGENT
    agents:
      - step: 03
        skill: dig-axioms
        input: "02-steelman.md"
        sub_turns:
          - "03a-level1.md"
          - "03b-level2.md"
          - "03c-level3.md"
        output: "03-dig-axioms.md"

  - turn: 4
    
        
         
         
         

      
     
    
        
         
         
        
             
                  
                    
         

    
    
        
         
           
         

    
    
        
         
         
         

Convergence criteria:

• All 7 turns complete
• FINAL-synthesis.md exists
• Optional: Reflection-loop includes confidence score ≥0.80

---

Level 4: Epic (5+ agents, 5-10 turns, BD-tracked)

Use case: Strategic analysis, multi-session work, unknown unknowns

Complexity factors: 10+

• Strategic decision
• Multi-session work
• Unknown unknowns
• BD epic required

Pattern: BD-tracked phases with dependencies

Example: Strategic Bet Analysis

Prompt: "Should we bet on technology X? Full strategic analysis."

workflow: strategic-bet-analysis
bd_epic: true  # Track in BD with dependencies
phases:
  - phase: research
    turns: 2
    bd_milestone: "clawd-xxx.1"
    agents:
      - skill: "hidden-gems/ergodicity"
      - skill: "hidden-gems/lindy-effect"
      - skill: "hidden-gems/cynefin"
      - skill: "domains/systems-thinking/*"
      - skill: "axioms/paraphrase-engineer"

  - phase: analysis
    turns: 3
    bd_milestone: "clawd-xxx.2"
    depends_on: research
    agents:
      - skill: "axioms/dig-axioms"  # 3-turn
      - skill: "axioms/multi-devils-debate"  # 3-turn
      - skill: "hidden-gems/morphological"
      - skill: "hidden-gems/triz-40"

  -  
     
     
     
    
        
        

    
     
     
     
    
        
        

BD Structure:

bd://clawd-xxx (Epic: Strategic Bet Analysis)
├── clawd-xxx.1 (Research phase) ✅
├── clawd-xxx.2 (Analysis phase) - blocked by clawd-xxx.1
├── clawd-xxx.3 (Synthesis phase) - blocked by clawd-xxx.2
└── clawd-xxx.4 (Output phase) - blocked by clawd-xxx.3

Convergence criteria:

• All 4 phases complete (BD milestones closed)
• Synthesis phase reaches convergence threshold (0.85)
• Output phase produces actionable recommendation

---

Convergence Criteria (CRITICAL)

When to Stop a Deep Workflow

Problem: Without clear exit criteria, Deep workflows can loop infinitely.

Solution: Define convergence criteria per workflow type.

---

Criteria Type 1: Turn Count (Fixed Structure)

Use for: Workflows with defined structure (e.g., axiom exploration)

Rule: Stop when all defined turns complete

Example:

workflow: axiom-exploration
turns: 7  # Fixed structure
convergence:
  type: turn_count
  required_turns: 7
  exit_condition: "All turns complete"

Validation: Check that output files exist for all 7 turns

---

Criteria Type 2: Confidence Threshold (Iterative)

Use for: Root cause analysis, hypothesis testing

Rule: Stop when confidence ≥ threshold OR max iterations reached

Example:

workflow: root-cause-analysis
convergence:
  type: confidence
  threshold: 0.85
  max_iterations: 5
  evaluator_skill: "evaluate-confidence"

Execution:

Turn 1: Generate hypotheses → confidence = 0.60
Turn 2: Test hypotheses → confidence = 0.75
Turn 3: Refine top hypothesis → confidence = 0.88 ✅ STOP

Safety: Always include max_iterations to prevent infinite loops

---

Criteria Type 3: Perspective Coverage (Fan-out)

Use for: Multi-perspective analysis (Base workflows)

Rule: Stop when all requested perspectives have been collected + synthesis complete

Example:

workflow: pros-cons-analysis
convergence:
  type: perspective_coverage
  required_perspectives: ["pros", "cons", "synthesis"]

Validation: Check that all perspective files exist (01a-pros.md, 01b-cons.md, 02-synthesis.md)

---

Criteria Type 4: Resonance Loop Convergence (Epic)

Use for: Strategic decisions requiring consensus

Rule: Stop when consecutive iterations show <10% change in output OR max iterations

Example:

workflow: strategic-decision
convergence:
  type: resonance
  max_iterations: 5
  convergence_threshold: 0.85
  delta_threshold: 0.10  # <10% change = converged

  loop:
    - skill: synthesize-current
    - skill: critique-synthesis
    - skill: integrate-critique
    - check: |
        delta < 0.10 AND confidence >= 0.85?
          yes: exit_loop
          no: continue (up to max_iterations)

Anti-pattern: Orchestrator synthesizing → Groupthink Pattern: Separate critique + integrate agents → No groupthink

---

Convergence Validation Checklist

Before declaring a Deep workflow complete, verify:

• Exit criteria met (turn count, confidence, perspective coverage, or resonance)
• All artifact files exist (no missing turn outputs)
• FINAL-synthesis.md exists (final step completed)
• Max iterations not exceeded (safety check)
• Confidence score documented (if applicable)

If criteria NOT met: Report blocker and either:

1. Add more turns (if under max_iterations)
2. Lower threshold (if unrealistic)
3. Escalate to human (if truly stuck)

---

Resonance Loops (When Needed)

For convergence on complex questions:

resonance:
  enabled: true
  max_iterations: 5
  convergence_threshold: 0.85
  check_skill: "evaluate-confidence"

  loop:
    - skill: synthesize-current
    - skill: critique-synthesis
    - skill: integrate-critique
    - check: confidence >= threshold?
      yes: exit_loop
      no: continue

Anti-pattern: Orchestrator synthesizing → Groupthink Pattern: Separate critique + integrate agents → No groupthink

---

Agent Dispatch Pattern

For each turn:
  If parallel:
    Launch all agents simultaneously (Task tool, multiple calls)
    Wait for all to complete
  Else:
    Launch sequentially

  Sync: Verify outputs exist before next turn

---

BD Integration (Epic Workflows)

Before Starting Epic Workflow

# Create BD epic
bd create --title="Workflow: <name>" --type=epic --priority=P1

# Example: Strategic bet analysis
bd create --title="Epic: Evaluate GraphQL migration" --type=epic
# Returns: clawd-xxx

Per Phase

# Create phase task
bd create --title="Phase: Research" --type=task --parent=clawd-xxx
# Returns: clawd-xxx.1

# Add dependency to previous phase
bd dep add clawd-xxx.2 clawd-xxx.1  # Analysis depends on Research

On Phase Complete

# Close phase milestone
bd close clawd-xxx.1

# Check if next phase unblocked
bd show clawd-xxx.2  # Should show: unblocked

On Workflow Complete

# Close epic
bd close clawd-xxx

# Archive artifacts (optional)
mv tmp/strategic-bet-* ~/.lev/pm/reports/clawd-xxx-strategic-bet.md

---

Anti-Patterns to Avoid

Don't	Do Instead
Orchestrator reads all context and synthesizes	Dispatch separate synthesis agent
Agents share context during execution	Each agent only reads assigned input files
Single mega-agent doing everything	Break into focused single-responsibility agents
Pseudo-code functions that can't execute	Reference real skills that exist
Over-engineer CDO "properties"	Graph structure + file I/O + agent dispatch
No max_iterations on loops	Always set safety limit (5-10 iterations)
Vague convergence criteria	Define explicit exit conditions

---

Relates

Depends On

• skill://lev-cdo/router - Provides classification and routing
• skill://lev-cdo/skill-discovery - Finds skills to attach to workflow nodes
• skill://bd - Tracks Epic workflows

Works With

• skill://planning - CDO outputs feed into planning
• skill://work - Validates CDO outputs via lifecycle ALIGN gate

Blocks/Enables

• Blocks: Execution if convergence criteria not met
• Enables: Multi-perspective analysis without groupthink

Lifecycle Position

• Before: Classification + routing decision
• After: FINAL-synthesis.md → proposal.md or report.md
• Parallel: None (sequential execution per turn)

Technique Map

• 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 workflows, 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.