AI code review before every pull request

The model that wrote a change is the worst judge of it — it already decided every questionable line was fine. claudrunner hands every diff to a fresh, hostile reviewer before a pull request opens.

Why a fresh reviewer

Asking the same session to review its own work produces a clean bill of health almost every time. The reviewer in claudrunner is a separate agent with no memory of writing the code. It receives the raw diff — never a summary, because a summary carries the author’s own framing, which is exactly what a review exists to escape.

Its job is to find what passing tests cannot: race conditions, a comparison that fires when nothing changed, a new status its consumers never handle, a query that grows with your data, a contract a client still depends on.

How findings are graded

Grade	Meaning	What happens
P0	Fatal at runtime: a crash, corrupted data, exposed data	Fixed before the pull request opens. If it cannot be fixed with confidence, that change is reverted and parked as `failed-review`
P1	Wrong behaviour under realistic conditions — silent, which is worse	Fixed before the pull request opens, and the affected tests re-run
P2	Minor	Listed in the pull request under Known minor findings, not fixed

Every finding needs a file, a line range, a one-sentence failure mode and a concrete reproduction. Anything without code evidence is not reported, and “no findings” is a valid answer — it never manufactures concerns.

A second opinion from another vendor

Optionally, a read-only command-line tool from a different model vendor reviews the same diff in parallel. Two vendors disagreeing is a much stronger signal than one model checking itself. If that tool fails, the run records the real error and carries on with the main reviewer — it never invents the missing review.

review:
  second_vendor:
    enabled: true
    command: "your-other-cli --read-only"

It gets better at your codebase

When the reviewer catches a failure mode that will recur, it is written into .claudrunner/gotchas.md with the evidence. Every future review reads that file first. Over a few weeks the reviewer becomes specifically good at your repository’s traps.

Use it on your own work too

The review is a skill you can call from any Claude Code session, not only inside scheduled runs. Say “review before PR” or “is this ready?” and it reviews your branch the same way.

How a run works — where the review sits in a run.
Nightly AI security scan