With a cap of 2 the second call is refused, so every build that uses its full budget fails at the last step.
A reviewer, not a linter.
Connect a repository and every pull request gets read against the whole codebase, not just the diff. It comments when something is actually wrong, and stays quiet the rest of the time.
Quiet by default
Most review bots fail by talking too much. Everything here is tuned so that a comment means something is wrong.
Nothing to say, nothing said
A clean pull request gets no comment at all. No "looks good", no summary of what you already know.
Nits stay unposted
Anything below medium severity is recorded and kept off the pull request. Ten comments is the ceiling, and you can lower it.
Push again, hear nothing twice
Each push is reviewed against what was already reported. The same finding is never raised a second time.
It reads the repository, not the diff
The review runs against a real checkout. When a change alters an assumption, it goes looking for the code that depended on it, which is where the expensive bugs live and exactly what a diff cannot show.
It reads the issues a pull request says it closes, so it can check the change against what it claims to do instead of guessing from the branch name.
- req.Header.Set("Idempotency-Key", key)
billing/refund.go:212 still expects the key to be set and retries on a timeout. Without it, a retried refund charges twice.
Findings that carry evidence
A review is a first pass, and your test suite finishes after it does. What CI learns gets attached to the finding it concerns.
The check that corroborated it is named. Recorded on the existing comment, with no second notification.
The finding said up front what would prove it wrong. That check passed, so the comment was withdrawn and the thread resolved, quietly.
A finding is only ever withdrawn by a check it named itself. A generally passing suite is never treated as proof, because most of the time it means nothing tested that behaviour.
Reviewed on your terms
Every repository gets its own settings, changeable whenever.
Enabled
Reviews every pull request
Severity shown
Lower findings are recorded, not posted
Tone
How findings are worded
Language
Prose only, code is never translated
Model
Reviewing is a judgment task
Comments per review
Worst first, the rest are counted
Committable suggestions
One-click fixes on the diff
PR walkthrough
Summary on the first review
Draft pull requests
Skipped until marked ready
Ignore paths
One glob per line
House rules
Layered on top, never a replacement
It comments. That is all.
Code review never writes code, never opens a pull request, and never merges anything. It reads a checkout of the commit under review and posts comments. Suggested fixes are yours to accept or ignore.
Access is granted per repository through a GitHub App you install, and you can revoke it from GitHub at any time.
Put it on one repository
Connect a repository, open a pull request, and read what comes back.