Before a large refactor, Claude Code must survey around 300 files to catalog every place a deprecated configuration API is used. In previous attempts, the discovery output filled the context window before the refactor even began, and the implementation phase suffered. How should you structure the work?

Answer: D — Delegate the discovery to the Explore subagent, which isolates the verbose search output and returns a concise summary of usages for the main session to act on.

The Explore subagent isolates verbose discovery output and returns a summary, which prevents context exhaustion during multi-phase tasks while the main session keeps the high-level coordination (tasks 3.4 and 5.4). Printing only file names (a) loses the usage details the refactor needs and still accumulates in the main context. An output token limit (b) doesn't change how much discovery output piles up in context. Reading all files upfront (c) is exactly what exhausts the context, instead of building understanding incrementally.