I was researching whether Claude Code Review covered what I wanted from an AI reviewer, including running code to check its findings. The useful distinction was what Anthropic means by verification in the managed service.
The Code Review documentation describes agents reading the diff and surrounding code, then checking candidate findings against that code. It explicitly says the service does this without building or running the code.
| Question | Managed Code Review behaviour |
|---|---|
| Does it check a finding? | A verification step checks it against the code. |
| Does it execute the project? | It doesn't build or run the code. |
| Do findings block a merge? | It never approves or blocks the PR, and its check run doesn't conclude as a failure. |
That last part matters if I want a review gate. GitHub's required status checks accept successful, skipped or neutral results. Requiring a completed review check doesn't, by itself, make its findings merge blockers. Approvals, conversation resolution and test checks are separate requirements to configure.
For my comparison, finding quality, execution evidence and merge enforcement need separate answers. A reviewer can produce useful comments without reproducing the bug or enforcing a decision about it.
This was documentation research, not a local benchmark. Whether the managed reviewer's findings are good enough for my own PRs is still open. Running a test to reproduce a finding would be a separate experiment.