co-maintainer, an imitator of yours

One agent writes the change. Another agent reviews it. Neither one has read your repository.

An AI reviewer sees a diff. A human reviewer sees a diff plus everything they learned here over the past years.

A missed finding is usually a missing input, not a weak model.

Context is built once

Before it reads a patch, it reads the repository. Then it writes down what it found.

The guides live under your config directory, not inside the clone, so the context survives a fresh checkout and can be shared with a team.

Then it reviews

On the next pull request, the guides are already there.

Fix handshake schema compatibility #4128 GroophyLifefor:fix/handshake Open Merged

Review in progressAll checks have passed 0 of 1 check1 successful check

review server webhook received · starting reviewcontext loaded · posting findings

Review / co-maintainer running…2m 14s

A green check is not the interesting part. What it wrote on the diff is.

What it wrote

The pull request adds a field. Codegraph finds nothing that reads it.

#4128 Files changed

src/remote/server.ts in this PR

+minClientSchema: 2

codegraph-callers minClientSchema

0 callers · only written in this patch

co-maintainer reviewed just now

The server now claims it can refuse old clients. Nothing in the graph reads this field, so nothing is refused.

And it writes the fix

The missing half of the handshake belongs on the client.

Suggested change src/remote/client.ts

+if (schema < hs.minClientSchema) upgrade()

serverwrites minClientSchema

clientnow reads minClientSchema

The question is not which model reviews best, the question is about context management.

Lower context bloat → lower hallucination → faster reviews → more accurate predictions.