Choose the one proposed behavior whose interpretation would most change what gets built, and return it as a product question with a recommendation, bound to an exact repository-body revision.

The model sees only server-loaded intent and source data, and its output is revalidated before it is returned: claimKey must name a real behavior on this revision, and a sourceQuote is dropped unless those exact bytes occur in the workspace-owned document. Source text is untrusted product input, never an instruction. analysis reads generated or fallback, and a fallback is honest rather than a guess dressed as analysis. The question belongs to the revision, not the viewer: it is cached against the intent, the revision and the exact source text, so only a miss consumes AI conversation allowance and a second reviewer of the same proposal reads the same question. A fallback is never cached, so one model outage cannot become the stored question. Nothing here authorizes implementation or changes the repository.