100monkeys-classic.yaml

# 100monkeys Classic Workflow
#
# This workflow replicates the original hardcoded ExecutionSupervisor behavior:
# Generate → Execute → Validate → Refine (loop)

apiVersion: 100monkeys.ai/v1
kind: Workflow

metadata:
  name: "100monkeys-classic"
  version: "1.0.0"
  description: "Classic iterative refinement loop (generate → execute → validate → refine)"
  labels:
    category: "iterative"
    pattern: "generate-execute-validate-refine"
    stability: "stable"

spec:
  # Global context available to all states
  context:
    max_iterations: 10
    validation_threshold: 0.95

  # Start with code generation
  initial_state: GENERATE

  states:
    # ========================================================================
    # State 1: GENERATE
    # Execute the agent to generate/refine code
    # ========================================================================
    GENERATE:
      kind: Agent
      
      # Agent ID will be passed as workflow parameter
      agent: "{{workflow.agent_id}}"
      
      # Input template with conditional refinement feedback
      input: |
        {{#if blackboard.iteration_number}}
        === ITERATION {{blackboard.iteration_number}} ===
        
        Your previous attempt failed validation.
        
        --- PREVIOUS CODE ---
        {{blackboard.previous_code}}
        
        --- VALIDATION ERRORS ---
        {{blackboard.validation_errors}}
        
        --- REFINEMENT INSTRUCTIONS ---
        Fix the code to pass validation. Focus on:
        {{blackboard.validation_reasoning}}
        
        {{else}}
        === INITIAL ATTEMPT ===
        
        {{workflow.task}}
        {{/if}}
      
      # Timeout for code generation
      # Always proceed to execution after generation
      transitions:
        - condition: always
          target: EXECUTE

    # ========================================================================
    # State 2: EXECUTE
    # Run the generated code
    # ========================================================================
    EXECUTE:
      kind: System
      
      # Command will be passed as workflow parameter (e.g., "python main.py")
      command: "{{workflow.command}}"
      
      # Pass generated code via environment variable
      env:
        CODE: "{{GENERATE.output}}"
        PYTHONPATH: "/workspace"
      
      workdir: "/workspace"
      # Branch based on execution result
      transitions:
        - condition: exit_code_zero
          target: VALIDATE
        
        - condition: exit_code_non_zero
          target: REFINE
          feedback: "Execution failed: {{EXECUTE.output.stderr}}"

    # ========================================================================
    # State 3: VALIDATE
    # Judge validates the execution output with gradient scoring
    # ========================================================================
    VALIDATE:
      kind: Agent
      
      # Use the basic-judge agent (gradient scoring 0.0-1.0)
      agent: "basic-judge"
      
      # Input for judge with task, code, and output
      input: |
        === VALIDATION REQUEST ===
        
        Original Task:
        {{workflow.task}}
        
        Generated Code:
        {{GENERATE.output}}
        
        Execution Output:
        {{EXECUTE.output.stdout}}
        
        {{#if EXECUTE.output.stderr}}
        Stderr:
        {{EXECUTE.output.stderr}}
        {{/if}}
        
        --- INSTRUCTIONS ---
        Evaluate the code and output against the original task.
        
        Use the gradient scoring rubric:
        - 1.0 (Perfect): Fully meets requirements, production-ready
        - 0.7-0.9 (Good): Minor improvements possible but acceptable
        - 0.3-0.7 (Needs Work): Significant issues requiring refinement
        - 0.0-0.3 (Poor): Critical failures, fundamental rework needed
        
        Respond with JSON matching your output schema:
        {
          "score": 0.0-1.0,
          "confidence": 0.0-1.0,
          "reasoning": "Detailed explanation of score",
          "suggestions": ["Specific improvement suggestions"],
          "verdict": "pass" | "refine" | "fail"
        }
      
      # Branch based on judge verdict (thresholds: pass >=0.70, refine 0.30-0.70, fail <=0.30)
      transitions:
        # Success: Verdict is "pass" (score >= 0.70)
        - condition: score_above
          threshold: 0.70
          target: COMPLETE
        
        # Refine: Verdict is "refine" (0.30 < score < 0.70)
        - condition: score_between
          min: 0.30
          max: 0.70
          target: REFINE
          feedback: "{{VALIDATE.output.reasoning}}"
        
        # Failure: Verdict is "fail" (score <= 0.30)
        - condition: score_below
          threshold: 0.30
          target: REFINE
          feedback: "Critical issues found: {{VALIDATE.output.reasoning}}"

    # ========================================================================
    # State 4: REFINE
    # Update blackboard with feedback and check iteration limit
    # ========================================================================
    REFINE:
      kind: System
      
      command: "update_blackboard"
      
      # Update blackboard state for next iteration
      env:
        ITERATION_NUMBER: "{{blackboard.iteration_number + 1}}"
        PREVIOUS_CODE: "{{GENERATE.output}}"
        VALIDATION_ERRORS: "{{state.feedback}}"
        VALIDATION_REASONING: "{{VALIDATE.output.reasoning}}"
      
      # Branch based on iteration count
      transitions:
        # Continue iterating if under max_iterations
        - condition: custom
          expression: "{{blackboard.iteration_number < workflow.context.max_iterations}}"
          target: GENERATE
        
        # Give up if max_iterations reached
        - condition: custom
          expression: "{{blackboard.iteration_number >= workflow.context.max_iterations}}"
          target: FAILED

    # ========================================================================
    # Terminal State: COMPLETE
    # Workflow succeeded
    # ========================================================================
    COMPLETE:
      kind: System
      
      command: "finalize_success"
      
      env:
        FINAL_CODE: "{{GENERATE.output}}"
        FINAL_OUTPUT: "{{EXECUTE.output.stdout}}"
        VALIDATION_SCORE: "{{VALIDATE.output.score}}"
        ITERATIONS: "{{blackboard.iteration_number}}"
      
      # No transitions = terminal state
      transitions: []

    # ========================================================================
    # Terminal State: FAILED
    # Workflow failed (max iterations reached)
    # ========================================================================
    FAILED:
      kind: System
      
      command: "finalize_failure"
      
      env:
        REASON: "Maximum iterations ({{workflow.context.max_iterations}}) reached without success"
        FINAL_CODE: "{{GENERATE.output}}"
        LAST_ERROR: "{{state.feedback}}"
        ITERATIONS: "{{blackboard.iteration_number}}"
      
      # No transitions = terminal state
      transitions: []