If Cursor Bugbot doesn't automatically review a PR as you expect, don't rush to think that the GitHub App is hanging. According to Cursor's current Bugbot documentation, there are three most common reasons for this situation: the target repository is not enabled at all, the PR is not created by yourself, or your personal settings have been changed to "only run when mentioned in comments".
The most misleading part of this feature is that when you see the documentation saying "every PR update is automatically reviewed", you naturally think that all PRs should be triggered automatically. In fact, the premise is that the repository is enabled in the Bugbot dashboard, and for personal accounts, Bugbot will only run PRs from your own author by default.
If you previously cut your personal settings to mention-only, the behavior will be more like "you have to comment on cursor review or bugbot run before you start". In this case, on the surface, it "suddenly stopped running automatically", but the root cause actually changed the configuration state.
The most practical troubleshooting sequence is:
1. Go to Cursor's Bugbot dashboard to see if the target repository is actually enabled.
2. Confirm that this PR is not created by yourself.
3. Check if your personal settings have been changed to only comment triggers.
4. Try cursor review directly in the PR when needed.
If the comment trigger can run, it basically means that the GitHub installation and basic permissions are not bad, and the problem is more likely to be in the automatic trigger condition rather than the overall link.
Another overlooked point is that team members often think that "the organization has a bugbot and all PRs will be swept away." The document actually clearly states the boundaries, it can be controlled by repository switches and personal modes, not by default, and the whole repository and full PR run automatically without any difference.
Therefore, Bugbot does not automatically review PRs, but first checks whether the trigger conditions are met, which is often more effective than checking the model and network first.
