DSH plugins for Code Review

Review plugins aim to make an agent a better reviewer: establishing a baseline of intended architecture, checking whether a change drifted from it, requiring evidence before a claim of success, and summarizing risk in a diff. They do not replace a human on a merge decision; they make the first pass harder to bluff.

Prefer plugins that produce a reviewable artifact — a written report, a failing check, an approval prompt — over ones that only assert the code looks fine. A review that cannot fail is not a review. The strongest listings in this category treat “done” as a claim that needs evidence, not a sentence the model is allowed to utter.

How to evaluate

Look for three mechanisms. Baseline: does the plugin record what the system was supposed to look like before the change? Drift: can it show what moved away from that baseline without you re-reading the whole tree? Gate: does the agent have to attach a check, a test result, or a human approval before a task is marked complete?

Also check scope. A review layer that only scores the model’s own summary is theatre. One that reads the diff, the failing tests, or an architecture map is closer to the job.

What good looks like

A strong review plugin is explicit about what it measured. It names the contracts or architecture nodes it compared. It refuses to stamp success when evidence is missing. It keeps a durable record you can open later — not only a chat message that scrolls away. Task boards and multi-agent orchestrators often sit next to this use case: they help when work is genuinely parallel, and add ceremony when the task is a one-file fix.

If two plugins overlap, prefer the one that can show a failing path. “The agent said it reviewed the code” is not a control.

Common failure modes

The common failure is a green badge driven by the same model that wrote the change. The second is a baseline that is never updated, so every later diff looks like drift. The third is reviewing only the last message instead of the working tree — the model’s narrative of the edit, not the edit itself.

Install order and boundaries

Prefer a code-review plugin before another general-purpose agent bundle. Make diffs and suggestions reviewable before you allow automated commits. Ungated auto-fix will change the repo while you are away.