Code Review

Beta

Agent-performed review grounded in repository conventions, invariants and history.

Review is the inverse of editing: instead of producing a change, the agent evaluates one. The same four-layer model applies, which is what separates useful review from restating a style guide.

What is actually checked#

Linters already catch formatting and a long tail of local mistakes. There is no value in an agent duplicating them. Review is scoped to the classes of defect that require understanding the repository as a whole.

ClassWhat it looks like
Invariant violationA change breaks a property the codebase relies on elsewhere but does not state locally
Incomplete blast radiusA signature changed in one place and in four of its five call sites
Convention divergenceThe change is correct but does not resemble how this repository solves the same problem elsewhere
Coverage regressionNew behaviour lands on a path no test exercises
Repeated defectThe change reintroduces a pattern that history shows was previously reverted or fixed

Conventions are derived, not configured#

Rather than asking a team to write down its conventions, the reviewer infers them from the repository and states them as observations with a confidence level. A pattern followed in forty of forty-two comparable sites is a convention; one followed in nine of twenty is not, and is not reported.

Review finding
finding    convention divergence          confidence  high
location   billing/invoice.ts  ·  createInvoice

  This module's 11 other write paths wrap their database
  calls in withTransaction(). This one does not.

  comparable sites   11 of 11 use withTransaction
  counterexamples    none
  consequence        a failure between the insert and the
                     ledger update leaves an invoice with
                     no corresponding entry

  suggested          wrap lines 40-58; no signature change

Scope limits#

The reviewer does not approve, merge, or gate. It produces findings; a human decides. This is not a temporary limitation pending better models — review authority is a human accountability boundary, and the runtime is built to keep it that way.

Last updated 2026-09-02