mexxi

0007. Implementers implement, reviewers judge ​

Status: Accepted · Date: 2026-10-21

Context ​

An agent that reviews its own work tends to approve it, and a reviewer that also fixes what it finds turns review into another round of implementation. Unbounded review loops cost time and tokens and rarely converge. When implementer and reviewer share a model, they also share its blind spots.

Decision ​

  • The code-reviewer agent judges and never edits. It works from the diff and a short background file with the requirements, not from the implementation transcript or the implementer's reasoning.
  • Review runs when independent judgement adds confidence: substantial behaviour, architecture, integration, security or user-facing changes. Small mechanical changes skip it.
  • There are at most two passes: discovery, then one confirmation that checks each blocker was resolved. A blocker found at confirmation stops the run and goes to the person.
  • The orchestrator checks every finding against repository evidence. Critical, high and medium findings block. Low findings inside the task's scope with a bounded fix are fixed; the rest are listed in the pull request.
  • Where the host allows it, the reviewer runs on a model from another family. Claude Code delegates through ibni cross-review to codex exec; OpenCode renders another model for the reviewer. Codex has one family, so its review is always same-family. The pull request's Evidence says which kind ran.
  • Every review ends with a merge-danger call: whether the change is a one-way or two-way door, and its blast radius.

Consequences ​

  • Review findings are evidence to investigate, never commands to obey.
  • A run always ends after two passes, so review can't loop indefinitely.
  • Cross-family review needs a second tool installed. Without it the review still runs, and the pull request says it was same-family.
  • The retro compares how many findings were accepted from same-family and cross-family reviews, so the extra cost has to show its value.

Alternatives considered ​

  • Self-review by the implementer: cheap, but it approves its own blind spots.
  • A reviewer that fixes its own findings: fewer hand-offs, but the review is no longer independent.