Nick Nisi of WorkOS presents a playful simulated fruit-fly code reviewer and a proposal for directing human attention to riskier changes.

Code smells, literally

  • A model turns properties of a code change into inputs for a simulated fruit-fly connectome; learned associations map those inputs to approach or avoidance.
  • In a three-example audience game, the fly got two right and the room got three. A subsequent real pull-request demonstration did not produce a clear decision.
  • Nisi reports roughly 88% accuracy in his own testing, but does not establish an independent benchmark or show enough evaluation detail to generalize it.
  • The output is a coarse accept/reject signal, not an explanation or a replacement for a conventional reviewer.

A proposed review pipeline

  • Maintain a team-specific review skill capturing the checks a team expects for its code.
  • Use a classifier and automated reviewers to assess complexity, risk and the change itself; escalate failures to a person.
  • Nisi proposes automatically merging changes that pass the checks, with incident review to improve the process. This is a proposal, not a demonstrated safety guarantee.
  • Give reviewers evidence such as screenshots, test results, command output and reproducible smoke tests, plus context about why the change was made.

“We need to understand why code is actually being changed, or why functionality is changing more specifically.”