19-refine-loop.vvm

# VVM Example 19: Refine Loop
# Iterative improvement using the refine helper
# This demonstrates the "Evaluator-Optimizer" pattern from Anthropic

agent coder(model="sonnet", prompt="Write clean, well-tested code.")
agent reviewer(model="opus", prompt="Review code critically. Find bugs and issues.")

# Code to review and improve
initial_code = """
def calculate_discount(price, discount_percent):
    return price - (price * discount_percent)
"""

# The evaluator: check if code passes review
def passes_review(code, iteration):
  review = @reviewer `Review this code. List any bugs, edge cases, or improvements needed.
If the code is production-ready with no issues, respond with "APPROVED".`(code)

  return ?`contains APPROVED with no significant issues listed`(review)

# The optimizer: improve based on feedback
def apply_feedback(code, iteration):
  # Get detailed review
  review = @reviewer `Review this code critically. Focus on:
1. Bug risks
2. Edge cases not handled
3. Missing input validation
4. Code clarity
Provide specific, actionable feedback.`(code)

  # Apply the feedback
  improved = @coder `Improve this code based on the review feedback.
Code: {code}
Review: {review}

Fix all issues mentioned. Add tests if appropriate.`(pack(code, review))

  return improved

# refine loops until passes_review returns true or max iterations reached
# Each iteration: evaluate -> if not done -> improve -> repeat
final_code = refine(
  initial_code,
  max=5,
  done=passes_review,
  step=apply_feedback
)

# The final code should handle edge cases, have validation, etc.
export final_code