You give a reviewer subagent the diff and a list of criteria, while the session that wrote the code stays open. Which account of that arrangement is right about what the reviewer can see, where its findings go, and what an open-ended brief will return?

The principle — a reviewer subagent has two properties that are easy to collapse into one, and a third thing that is entirely up to you: what it sees, where its output goes, and what you asked it for.

Why the key is correct — the first property is what it sees. A reviewer running in a fresh subagent context sees only the diff and the criteria you give it, not the reasoning that produced the change, so it evaluates the result on its own terms. The second is independent of the first and concerns routing: because the reviewer runs as a subagent, the implementing session receives the gaps directly and can fix them and re-review without you copying findings between windows — and the subagents guidance describes the same shape as a reviewer subagent that dispatches a verifier per finding, so the intermediate output never reaches your main conversation. You get the gaps; you do not get the reviewer's working. The third is the brief. A reviewer prompted to find gaps will usually report some, even when the work is sound, because that is what it was asked to do, so an open-ended brief returns findings on sound work by design.

Why the others are wrong — treating the absent reasoning as a weakness inverts the mechanism that makes the verdict worth having. Relaying gaps by hand describes the arrangement this one replaces, and it also misreads why your context stays clean — that follows from the subagent boundary, not from your filtering. Reading an open-ended brief as strictness is the trap the guidance names: the volume reflects the ask, and chasing all of it produces extra abstraction layers, defensive code, and tests for cases that cannot happen. The remedy is to flag only gaps that affect correctness or the stated requirements and treat the rest as optional.

Remember this — the reviewer's isolation buys you an unbiased verdict and a clean context; the usefulness of what it says is bought separately, by the brief.

Sources — Anthropic's Claude Code best practices and subagents documentation.