Parse alternate reasoning fields, retry with a larger max_tokens budget, cap review images, and fall back to text-only when vision returns empty. Co-authored-by: Cursor <cursoragent@cursor.com>
dekun