How to review AI-generated pull requests: check what the change affects.
Pull requests written by Claude Code, Codex or Cursor read well. When they break something, it is usually code the diff doesn't show. Veridu looks for that code and reviews the change with the context it finds.
- Callers it didn't update. The agent changed a signature or a default; code elsewhere still uses the old one.
- Subclasses and implementations. A new call or a changed contract no longer fits a class the pull request never touched.
- References to removed code. Something was renamed or deleted, and something still calls it.
- Documented behaviour. The behaviour changed; the docs, and the apps relying on them, still describe the old one.
- Fixes applied in one place only. The agent fixed the pattern where it was looking and left the same code unchanged elsewhere.
Two real examples, from Laravel and Fastify: pull request impact analysis. To check by hand: how to find what a change breaks. Worried about noise: how Veridu decides what can block.
- 1Install the Veridu GitHub App and link the repository. Link each repository on the Repos page on veridu.io.
- 2Open the pull request ready for review. Veridu reviews it once, automatically, when it is opened ready or marked ready, and posts findings as a PR comment and a Veridu check. Pull requests opened by bots such as Dependabot are skipped. Drafts are not reviewed unless the repository has Review before a human turned on.
- 3Read what it found outside the diff. Each finding names the file, line and evidence. Findings that cite a file or line that does not exist are dropped; the rest still need your judgement. A problem blocks the merge only when reviewers find it independently and rate it at or above your blocking threshold.
- 4Push fixes and ask again. Unless Review before a human is on, new commits are not reviewed automatically: the check turns neutral. Comment /veridu review on the pull request to review the current commit.
The agent can check its own change before it commits, so fewer findings reach the pull request: in Claude Code or in Codex. To have Veridu review every push to a draft before a person looks at it, see review before a human.
Why review AI-generated pull requests differently?
An agent's diff is usually tidy and reads as correct. What it misses tends to sit outside the lines it touched: a caller it did not update, a subclass it did not know about, a doc page describing the old behaviour. Reading the diff more carefully does not show you those.
Which languages are traced automatically?
Callers, subclasses and implementations are traced before review for PHP. For TypeScript, JavaScript and Python, Veridu detects changed definitions, and code a fix left unchanged elsewhere, and fetches the code that uses them. Other languages and files such as YAML, Terraform or Dockerfiles go to reviewers that read and search the checkout themselves.
Will it block my merges?
Only when reviewers find the same problem independently and rate it at or above your blocking threshold. Other findings are left for the human reviewer; those below the repository's display level can be left out of the PR comment. A review that could not see all of the changed code says so and reads as review, never pass.
Does it approve pull requests?
Never automatically, and never as the GitHub App. It reports what it checked and what it found. An eligible person with a linked personal token can submit their own GitHub approval from the Veridu review page.
Can I try it without installing the app?
Yes, on a public pull request. Add Veridu to Claude Code or Codex and ask it to review the pull request URL; it calls veridu_review.
What does it cost?
Up to 20 reviews a month on public GitHub repositories are free, no card. Solo and Team plans add private repositories.
Free for public repositories, up to 20 reviews a month. See pricing for private repositories.