Gito

The deciding trait is that it reads your uncommitted changes. Most review bots wake up when you open a pull request, so the first person to see your mistake is a colleague. This one runs against the working tree, which means the first pass happens on your machine and nobody watches you take the note.

The rest is unremarkable in the good sense: no signup, no dashboard, no seat to buy. Pick it if your team is small and the review queue is where changes go to wait. Pick CodeRabbit if you would rather pay somebody to keep it working.

The failure mode of a review agent is noise, and this one has no documented dial. No severity threshold, no per-category suppression, and no configuration for how much it may say about a single change. A reviewer that comments on everything gets trained out of a team within a fortnight.

It also stops at the comment. It writes the finding and never commits or opens a branch, so every accepted suggestion is still work somebody does by hand afterwards. What it does right is refusing to touch the branch at all. A review agent that also commits is a review agent nobody reads.

1. Provider agnosticism carries an epistemic cost that is rarely acknowledged. If the reviewer is whichever model the operator wired in, precision and recall become properties of a deployment rather than of the software, and two teams running this are not running the same reviewer at all.

2. The correct response would be a fixed reference configuration with a measured error rate against it, so a reader has one number to argue with. 3. None exists. The design is defensible and unmeasured, and for a review agent the measurement is the product, which makes the omission the interesting part.

The category is the problem, not the code. Review automation is one of the few places in this space where buyers reliably pay, which has attracted companies with sales teams and dashboards and a reason to keep shipping. Competing there from a volunteer repository is possible and it is not a business.

Moat: none, unless the habit forms before somebody bundles the same feature into a platform teams already pay for. Likely path is steady maintenance at small scale, or quiet abandonment. Position: use it, contribute a fix when you need one, and do not expect a support queue to answer you.

This is the rare one that fits where the work already happens. It installs as a job on the pull request, so nobody learns a new tool and nothing changes about how a change gets merged. It also reaches the trackers my teams run, so a finding lands where the work is tracked.

The bill is inference on every pull request, which is a number I can model from last quarter's volume and then cap. There is no seat price and no vendor contract, and therefore no vendor to hold to anything. Approved with conditions: one repository first, and a spend cap before it goes wider.

The licence is the permissive kind and the install is a single package with no account, no daemon and no service to register with. I like tools that go in with one line and come out with one line, because that is the difference between trying something and committing to it.

What it does not have is a way into anything else. No protocol server, so my own agent cannot call it as a tool, and no local runtime is documented, so the review leaves the building even when the diff does not. Grudging respect for the install and a shrug at everything after it.