Skip to main content

What is supported

Last updated on

This page lists what AI Code Review supports today. Support changes during beta, so check this page before you file an issue.


Source control platforms

PlatformStatusNotes
Harness Code RepositorySupportedNo connector required. Enabled by feature flag, then per repository.
GitHub CloudSupportedUses a standard Harness GitHub connector.
GitHub Enterprise ServerSupportedSame connector model as GitHub Cloud.
GitLabComing soonNot available today.
BitbucketComing soonNot available today.

GitHub repositories are reviewed through a linked Harness Code repository rather than in place. This is why GitHub onboarding waits for an import to finish and Harness Code onboarding does not.


Connectors

AI Code Review uses standard Harness GitHub connectors. There is no separate AI Code Review GitHub App to install, and most teams already have a suitable connector.

PropertySupported
AuthenticationAny type the Harness GitHub connector supports, including personal access token and OAuth
Connector scopeAccount-level or repository-level
Provider repository identifierRequired in owner/repo form for account-level connectors, must be empty for repository-level connectors

Go to GitHub integration to review the connector requirements in detail.


Models

ProviderModelsNotes
AnthropicAny Anthropic model, including Sonnet, Opus, and FableOpus is recommended
Anthropic on Amazon BedrockAny Anthropic model available in your Bedrock accountFor accounts that require models to stay inside AWS

Anthropic is the only provider supported at beta. Bring your own key, or use the Harness AI gateway.


Configuration scopes

A setting belongs to exactly one space or exactly one repository, never both.

ScopePath formApplies to
AccountaccountEvery project in the account
Organizationaccount/orgEvery project in the organization
Projectaccount/org/projectEvery repository in the project
Repositoryaccount/org/project/repoThat repository only

A space path has at most three segments. A path with four or more segments is rejected.


Review triggers

EventReviewed
Pull request createdYes
Pull request updated with a new commitYes
Pull request reopenedYes
Pull request closed or mergedNo
Comment or label addedNo
Manual re-runThrough the pipeline, not through the pull request

A new push aborts a review still running for the same pull request, so only the latest commit is reviewed.

There is no branch, author, or label filtering on the trigger. Every pull request in an onboarded repository is reviewed.


Result states

StateMeaning
successThe criterion passed.
failureThe criterion did not pass.
errorNo verdict was produced. Pending and running checks also report as error through the API.

There is no pending or running result state in the API. If you are polling for completion, treat error as inconclusive rather than as a failure.


Risk levels

RiskAppears as
lowRisk: LOW
mediumRisk: MEDIUM
highRisk: HIGH

While a review is running, no risk level is shown.


Known boundaries

None of the following are defects. They are the shape of the beta, and each one has a practical consequence.

  • No hard limits. Diff size, repository size, review concurrency, and rate limiting are unbounded. Go to Limits and quotas to understand what that means in practice.
  • One check per criterion. There is no aggregate check to require in branch protection. Requiring AI Code Review on a merge means requiring the individual checks.
  • Criteria are not deduplicated across scopes. A criterion defined at both project and repository level produces two checks.
  • A setting cannot be locked against override. An account administrator cannot prevent a lower scope changing an inherited value.
  • Criteria cannot be stored in Git. There is no configuration-as-code path for review criteria, the connector, or the system prompt.
  • The agent cannot push. It has no push permission, so it cannot commit a fix.
  • A Harness account is required to see full results. Comments and checks are visible on GitHub without one, but the detail page is not.
  • Offboarding is partial. Removing a repository deletes its trigger and disables review on it. The pipeline, service account, role, token, secret, and stored settings remain.
  • Results are per commit. A force push past the reviewed commit removes the results from view.