Starter
Try AI PR review.
- 2 reviewers
- Code Review
- Security Review
- 3 seats included
- 15,000 credits to start, then 5,000/mo
- About 70 reviews to start, then about 25 a month
- Unlimited repos
Code review is a second engineer reading a change before it merges, to catch bugs and security holes before users hit them.
HighKnown issue in the Code Review benchmark · caught by DeepSeek V4.1 FlashSee the graded case
The most common miss sits outside the diff. Two AI readers independently graded 267 IonWarp findings, nearly all from 24 to 29 September 2026. Of the 17 that were real merge blockers, 12 were a test, a migration, a config value or a second read path the pull request forgot to change.
“Pull requests are proposals to merge code changes into a project.”[1] You push a branch, open one against main and request a code review under Reviewers in its sidebar;[2] the team reviews it there. What is a pull request covers forks, approvals and merge methods. Projects that work by email, such as the Linux kernel, ask for a merge with a message that git request-pull writes. We also publish a comparison of code review tools, from Gerrit to static analyzers and AI reviewers, at bestcodereviewtools.com.
Conversation holds the description and reviews; Commits, the branch's history; Checks, the tests and builds; Files changed, the diff the reviewer reads.[1]
A draft cannot be merged, and code owners are not asked to review it until you mark it ready for review.[1]
Know what the change is meant to do before you read the diff. The author's side is the pull request checklist.
A test covers the new behavior and would fail without the change. On a UI change, try the keyboard too: accessibility testing.
Search for the callers, tests, migrations, configs and docs of anything renamed, removed or retyped. In that read, one of our own pull requests took a release gate to 36 migrations and left a test expecting 35. Another hid a kind of run from the leaderboard and events, and the public runs endpoint kept serving it. IonWarp's reviewers raised all 17, so the sample leans toward what they look for. An AI reviewer can take this first pass: how the models scored.
Code ReviewOn Starter, Pro and Max
Errors are handled, never swallowed, and a retry cannot charge or send twice.
New routes check who is calling, input is validated, and no secret lands in code or logs. Security code review goes further.
Security ReviewOn Starter, Pro and Max
Mark what must change before merge, label what is optional as a nit, and approve once it has. Code review terms lists the shorthand: LGTM, PTAL, SGTM.
IonWarp's column is Code Review on a GitHub pull request. The last row is not a ranking: each figure was measured a different way.
| Criterion | IonWarp | Pull request review | Pair programming | Formal inspection |
|---|---|---|---|---|
| Who reads the code | An AI model, which reads the pull request and the code around it | A teammate, in a tool that shows the diff and holds the comments[3] | A second developer, the navigator, while the driver writes[3] | Several developers, line by line, in a series of meetings[4] |
| When it happens | On every pull request, and again when you push a commit | After the change is written, before it merges[3] | While the code is being written[3] | In multiple formal phases[4] |
| What has been measured | Code Review found 77 of 158 known issues on 50 benchmark pull requests. Qodo found the most, 98.[5] | About 15% of reviewers' comments at Microsoft indicated a possible defect, and at least 50% concerned long-term maintainability.[6] | In a 1999 University of Utah class experiment, pairs took about 15% more time and wrote code with about 15% fewer defects.[7] | Capers Jones's analysis of over 12,000 projects put its latent defect discovery rate at 60 to 65%.[4] |
One workspace price with seats included. Credits pool across your team. Unlimited repos on every plan.
Try AI PR review.
Get AI PR review right from day 1.
Scale AI PR review.
Custom reviewers + credits · GitHub Enterprise · SSO/SAML + audit logs · Custom security review + SLA
To catch defects, security problems and design issues before they merge, and to spread knowledge of the codebase across the team.
Keep pull requests small: Google's engineering practices call 100 lines a reasonable size and 1,000 usually too large, and set one business day as the longest a review request should wait for a response. SmartBear advises reading no more than 200 to 400 lines at a time (Google and SmartBear guides, Nov 2026). Read the description before the diff, comment on the code rather than the author, and separate blocking issues from suggestions.
Read the description first, then check the tests, follow anything renamed or removed outside the diff, look at the failure paths, check access and secrets, and mark what must change before merge. On GitHub you comment on the Files changed tab, then approve or request changes.
A pull request is the proposed change. The code review is the reading and discussion of it before it merges.
An AI model reads the pull request and the code around it and posts findings before a human reviews it. IonWarp runs Code Review on every pull request and adds specialist reviewers when a change needs them.
No. It takes the first pass (defects, security, duplication) so the human reviewer can spend their time on intent, design and product decisions. On 50 pull requests with 158 known issues, no AI reviewer we ran found more than 98 (IonWarp benchmark, Oct 2026).
isinstance check at line 254 tests against multiprocessing.Process which is the default context's ForkProcess class, not the SpawnProcess class created by multiprocessing.get_context('spawn'), causing the check to always be False on Linux and preventing process.kill() from being called on unhealthy workers · Code Review