Structure an automated PR reviewer as a scout→retrieve→lead pipeline instead of one big model call

Critique·medium signal

The emerging multi-agent code-review architecture runs six stages: webhook intake, a small fast 'scout' model that decides what in the diff matters, hybrid retrieval that pulls related code via semantic embeddings plus a symbol graph, then a frontier 'lead' model that produces structured findings and gates the merge. This scout-first design cuts frontier-model spend on trivial diffs and gives the lead reviewer the surrounding code it needs — the pattern behind reviewers that run 9 parallel sub-agents to catch bugs single-pass tools miss.