Limits and quotas
What AI Code Review does and does not bound at beta, and what that means for large pull requests.
AI Code Review ships no hard limits. Diff size, repository size, review concurrency, and rate limiting are all unbounded.
That is deliberate, not an oversight, and it has a practical consequence worth understanding before you turn it on across an organization.
What is not limited
None of the following are bounded:
Pull request or diff size. No threshold skips or rejects a review.
Repository size.
Review concurrency across repositories.
Number of review criteria per scope.
Criterion title and description length.
Number of MCP servers or tools per server.
Number of repositories onboarded in one request.
No limit is not the same as no ceiling
A very large pull request is not rejected by AI Code Review, but it can still fail at the model or in the pipeline. When that happens the failure surfaces as a criterion in the error state rather than as a clear message about size.
If large pull requests fail inconsistently while small ones succeed, treat size as the first hypothesis rather than looking for a limit to raise.
Enforced limits
Only two constraints are enforced, and neither is a quota:
Status check identifier length
127 characters
The criterion title portion is truncated silently. The criterion id is appended after truncation, so checks stay distinct.
Space path depth
3 segments
A path with four or more segments is rejected.
The repository listing API pages at 30 results by default and 100 at most, but that is pagination rather than a limit on what you can configure.
Concurrency within one pull request
One review runs per pull request at a time. A new commit aborts the review still running and starts a fresh one, so only the latest commit is reviewed.
Upstream limits that still apply
AI Code Review applies no rate limiting of its own. Two limits outside it still bite:
Model capacity. Sustained review volume across many repositories is bounded by your account model capacity rather than by AI Code Review.
Provider API limits. On the source control side, the status comment is best effort and is dropped rather than retried, so a provider rejecting it means the comment does not appear while the checks and findings still do.
Cost
Reviews consume model tokens on every run. With no hard limits at beta, cost is the real constraint, and it scales with where a criterion is defined rather than with how often it finds something: a criterion at organization level runs on every pull request in the organization whether it fires or not.
Go to Review criteria best practices to retire criteria that never fire.
Related concepts
What is supported: Platforms, models, and known boundaries.
FAQs: Common failures and questions.
Connectors and models: Model choice and cost.
Last updated
Was this helpful?