prcodereview Guides · by the Mergestorm team Try it free on a PR

Guides · Practice

What AI code review should flag on a pull request

Updated 29 Sep 2026 · ~6 min

Written by the Mergestorm team.

You have an AI reviewer on your repo, or you are about to. The useful question is not “does it use AI?” It is what it should leave on this diff, every time — and what it should keep to itself. If you want the definition first, start with what AI PR review is.

Three jobs a useful review does

An AI reviewer reads the change set, leaves feedback where a human reviewer would, and (if the product supports it) pushes a fix instead of only complaining. A useful one does three jobs well:

  1. Comments on the lines that matter — inline, not a vague summary at the bottom.
  2. Stays quiet when the change is fine — noise kills trust faster than a miss.
  3. Goes deeper when risk spikes — auth, money, migrations, wide refactors.

What you should see on a normal PR

If your AI reviewer only leaves a paragraph of praise and zero line comments, you bought a summarizer, not a reviewer.

How Mergestorm splits the work

On Mergestorm, the review is not one blob:

Maelstrom is a plan name, not an agent. Pick agents for intensity; pick the plan for billing.

How to roll it out without drowning

  1. Turn it on for every PR — selective coverage defeats the point.
  2. Require a bot pass before human review so seniors see a cleaner diff.
  3. Keep humans for product fit, architecture, and merge ownership.
  4. Keep PRs small (stacks help) so both the AI and the human stay sharp.

For the human half of the loop, use the PR review checklist (2026). For one agent vs a specialist fleet vs sticky follow-ups, see how to structure an AI PR review loop.

Quick takeaway

A reviewer that earns its keep comments inline on real problems, stays quiet on clean diffs, escalates on risk, and leaves the approve button to people who know the product.