The correct option is The extension sends the active editor content to a managed model endpoint in the cloud, the service computes a response based on that context and does not retain the submitted input after generating the suggestion.

This reflects how most cloud assisted coding tools work because the editor gathers relevant context from the active buffer and sometimes nearby files, sends a prompt to a hosted model endpoint, receives a completion, and then discards the transient request data. Providers commonly emphasize ephemeral processing for prompts and responses and state that prompts are not used to train the service and are not retained beyond what is necessary to deliver the suggestion.

All processing happens on the developer machine using a preloaded model and a local cache without any requests to cloud services is incorrect because that describes a fully local model which is not the typical pattern for a cloud backed suggestion service where inference happens on a remote endpoint.

Prompts are published to Cloud Pub/Sub and archived in BigQuery before a model creates a completion is incorrect because real time completions require a direct request to a model endpoint and archiving prompts in a messaging system or a data warehouse would be unnecessary and would raise avoidable privacy concerns.

The tool batches and stores every prompt locally then replays data from prior sessions to refine future suggestions is incorrect because retaining and replaying all prompts would conflict with privacy and least data principles and it would not be required to generate high quality suggestions in real time.

Full AWS Practitioner Certification Question