Skip to main content
A review belongs to a specific revision of a pull request. A newer commit changes the code under review, so evidence from an older run must not be read as coverage of the latest revision.

​
1. Read the change

The GitHub integration verifies the webhook and resolves the immutable PR revision. UI-impact detection considers source and contract changes that can affect the user experience, alongside the repository’s configured trigger rules.

​
2. Plan important flows

The planner combines change context with routes, objectives, checkpoints, and existing end-to-end coverage. A flow describes what a user should be able to do; its steps give the browser a concrete sequence to execute. Use repository configuration to describe important journeys and expected outcomes. A configured route is context for planning, not proof that it was covered. Check the executed plan and target evidence.

​
3. Execute the target matrix

Production runs the application in an isolated environment. Stagehand provides intelligent exploration and recovery; Playwright executes the recorded flow across the selected browsers and viewports. Every target needs its own recording. For example, selecting Chromium and Firefox with desktop and mobile viewports creates four browser/viewport targets. A completed desktop run does not substitute for an unfinished mobile run.

​
4. Analyze and reproduce

The review evaluates captured evidence for aberrations. Affected flows can be repeated to corroborate a finding. A failed attempt, missing artifact, or inconclusive reproduction stays explicit in the result. Repository rules and references ground brand findings. Reviewers should be able to trace a claim back to the behavior, expectation, and recording that support it.

​
5. Read the result

The production path updates the GitHub Check and summary with authenticated evidence links. Read both the findings and the coverage statement before relying on a result. Incomplete requested coverage cannot produce a successful Check. The local workflow uses Codex planning and deterministic analysis, stores runs locally, and does not publish to GitHub. Production exploration and multimodal analysis are not part of that local path.