How do you prevent coding assistants from reading irrelevant repo files?

Whenever I start a session in a large monorepo the agent automatically indexes or pulls in context from unrelated microservices. Suddenly, it starts suggesting patterns or types from a completely different service that shouldn't even be communicating with the current module.

The broader the context window gets, the more "cross-contamination" happens in the suggestions. How do you effectively curate or prune the file context before handing off a task to an agent?