In your code assistant's chat, a user working in a monorepo of about 30M tokens asks why payment retries double-charge. The model has a 200K-token window, and answer quality and latency both matter. How should context be assembled?

1. A Feed the entire repository across several turns in chunks so the model builds a complete picture of the codebase before answering.
2. B Rank files using open tabs, the symbol and call graph, and embedding search, then pack only the top results into a token budget.
3. C Include every file under the current directory recursively, since colocated code is most relevant and this avoids retrieval misses.
4. D Fill the full 200K-token window with the highest-similarity chunks, since supplying more relevant context always improves accuracy.