🔄 Data Flows & Sequence Diagrams

This document visualizes how request vectors, agent states, and evaluation variables flow across the AI Workflow Orchestrator.

Using formal Mermaid sequence flows, we break down three core operational patterns: Standard execution consensus, Security Veto intercepts, and Autonomous Self-Healing loops.

---

🟢 1. Standard Execution Flow (Consensus Proceed)

The sequence map below demonstrates a standard execution pipeline where the adversarial arena resolves with high consensus confidence and no security anomalies:

sequenceDiagram
    autonumber
    actor User as User / CLI / API Endpoint
    participant Engine as WorkflowOrchestrator
    participant Memory as Memory Controller (SQLite/Mongo)
    participant Arena as DebateManager
    participant Agg as DebateAggregator
    participant Exec as ExecutionManager
    participant Val as ValidationChecker

    User->>Engine: Submit Request (e.g. "deploy secure web app")
    Engine->>Engine: Assert Workspace (IdentityGuard)
    Engine->>Memory: load_memory_context() & retrieve_similar_workflows()
    Memory-->>Engine: Historic traces and agent Elo rankings
    Engine->>Engine: build_initial_plan()
    
    Engine->>Arena: Trigger Arena (run_debate)
    Note over Arena: Phase 1-5: Turn-based confrontation
    Arena->>Arena: 1. Analyst: Decomposes requirements & maps threats
    Arena->>Arena: 2. Solution: Drafts technical proposal plan
    Arena->>Arena: 3. Critic: Audits proposal and challenges flaws
    Arena->>Arena: 4. Security: Asserts sandbox permissions & bounds
    Arena->>Arena: 5. Optimizer: Recommends design refinements
    Note over Arena: Multi-Pass Heuristics ("Rule of 3" refinements)
    Arena-->>Engine: Structured arguments with confidence scores
    
    Engine->>Agg: Compute Consensus (calculate_consensus)
    Note over Agg: Weighting role priority + Elo ratings
    Agg->>Agg: Run Critic vs Solution Elo Duel (adjust ELO)
    Agg-->>Engine: Outcome (PROCEED: Consensus Reached, ELO updates)
    
    Engine->>Exec: execute_plan(final_decision)
    Exec->>Exec: Run sandboxed tasks (assert Admission Policies)
    Exec-->>Engine: Execution Report (Success)
    
    Engine->>Val: validate_results(execution_report)
    Val-->>Engine: Validation Report (Valid)
    
    Engine->>Memory: store_all_in_memory(outcome, trace, ELO)
    Memory-->>Engine: Commit confirmation
    Engine-->>User: Success response + Session Trace ID

---

🔴 2. Security Veto Intercept (Security Veto Triggered)

If the Security agent identifies a critical runtime privilege violation (such as mounting /var/run/docker.sock or hostPath volumes) with a confidence score $\ge 0.9$, the debate is immediately halted, execution plans are discarded, and penalties are applied:

sequenceDiagram
    autonumber
    actor User as User / CLI / API Endpoint
    participant Engine as WorkflowOrchestrator
    participant Arena as DebateManager
    participant Agg as DebateAggregator
    participant Memory as Memory Controller

    User->>Engine: Submit Request ("run container mounting docker.sock")
    Engine->>Arena: Trigger Arena (run_debate)
    Arena->>Arena: Solution Agent proposes Docker socket mount
    Arena->>Arena: Security Agent detects Privilege Escalation (Confidence 0.95!)
    Arena-->>Engine: Debate arguments
    
    Engine->>Agg: Compute Consensus (calculate_consensus)
    Note over Agg: Security confidence >= 0.9 forces immediate VETO
    Agg->>Agg: Penalty duel: Solution ELO -50 points
    Agg->>Agg: Reward: Security ELO +25 points
    Agg-->>Engine: Outcome (REJECTED: Critical security risk detected)
    
    Engine->>Memory: store_all_in_memory()
    Engine->>Engine: Audit Log Final Decision (trace_id)
    Engine-->>User: Error response: Access denied under GKE admission rules!

---

🟡 3. Autonomous Self-Healing Loop

If post-execution verification fails (e.g. system reports syntax issues or port blockages), the orchestrator intercepts the error and routes the state to the Self-Healing Loop to automatically patch the plan:

sequenceDiagram
    autonumber
    participant Engine as WorkflowOrchestrator
    participant Exec as ExecutionManager
    participant Val as ValidationChecker
    participant Heal as SelfHealHook
    participant Critic as Critic Agent
    participant Sol as Solution Agent

    Engine->>Exec: Execute Plan
    Exec-->>Engine: Execution failure (Error: Port 80 already in use)
    
    Engine->>Val: validate_results()
    Val-->>Engine: Validation Report (Invalid - failures detected)
    
    Note over Engine: Intercept validation failure!
    Engine->>Heal: attempt_self_healing(execution_report)
    Heal->>Critic: Diagnose Failure (think: "Explain port 80 fail")
    Critic-->>Heal: Diagnosis (Port 80 occupied, switch to port 8080)
    
    Heal->>Sol: Propose Repair Patch (think: "Generate repaired proposal")
    Sol-->>Heal: Repair Plan ("PROCEED: run command on port 8080")
    
    Heal->>Exec: Re-execute patch plan
    Exec-->>Heal: New execution report (Success)
    
    Heal->>Val: Re-validate results
    Val-->>Heal: Validation Report (Valid)
    Heal-->>Engine: Self-healing complete!