Yes, Opus 4.6 can refactor multi-file codebases safely when you scope changes, validate them with tooling, and ensure the model has access to the relevant files and build context. The model can propose coherent edits across files—updating interfaces and call sites, adjusting tests, and keeping naming consistent—but safety depends on your workflow, not just the model.
A safe refactor loop is: ask for a file-level plan, require a patch/diff output, apply it in a sandbox branch, then run lint/typecheck/tests and feed failures back for iteration. Keep changes small and incremental: migrate one module at a time, preserve public APIs unless explicitly changing them, and require a green main branch after each step. Also consider policy constraints like “no new dependencies” or “keep compatibility with a specific runtime version,” and bake those into prompts and CI gates.
Retrieval can reduce “project mismatch” errors. Store your coding standards, architectural patterns, and internal APIs in Milvus or managed Zilliz Cloud, retrieve the relevant conventions for the part of the repo being refactored, and include them in the prompt. This keeps refactors aligned with how your team actually writes code instead of generic patterns that look fine but fail review.