After gathering test results, the agent now pauses to audit its own work before writing the review: did it actually verify the highest-risk changes? Are there changed surfaces with no test at all? Did any test reach the right outcome through the wrong path? When it finds a fixable gap, it creates or edits tests and re-runs them before finalizing.
When a setup blocker prevents it from testing part of the PR – missing account credits, a feature flag left off, no required configuration – it now surfaces this under a collapsible "Suggestions" section with a "Fix in chat" link so the coverage isn't quietly dropped.