How do you make your code resilient to features in Machine learning model?

Current Model: Model A New Model: Model B

Example 1: If Model A has 10 feature inputs. However, Model B has only 6 feature inputs. No new features added. You have to make production code change remove using other 4 features in Production which requires a code change. How to make system intelligent to take only 6 inputs.

Example 2: If Model B doesnt workout well, you rollback to Model A. How do you ensure that all features are used again in Model A without any code changes and re-adding features.

How do you ensure that code is always compatible no matter what model you use in production.