Skip to main content
A review is tied to an immutable source: a PR revision, exact commit, or validated local-source snapshot. Another commit or changed local file requires another run.

​
1. Admit the request and resolve source

Diffraction checks workspace and repository access, review eligibility, organization pause and credit controls, and provider readiness. Automatic GitHub deliveries are deduplicated. Manual requests have their own request identity; retrying the same request does not intentionally create another paid review. Preparation resolves application configuration and project instructions from the selected source. PR network permissions come from its trusted base revision. Review each commit retains accepted revisions in a queue. Credit reservation happens at execution admission.

​
2. Select coverage and plan flows

UI-impact detection considers source and contract changes, trigger rules, and bounded agent classification where needed. The planner combines the change with routes, objectives, checkpoints, repository instructions, and private workspace context. Company rules and enabled checks are captured as immutable run settings. Selected native Playwright flows provide standing coverage before AI classification and planning. Native test outcomes remain separate from the AI review. A configured route or discovered test is context, not proof that it ran.

​
3. Prepare and execute

In isolated hosted execution, the application runs install, setup, build, start, readiness, and verification commands. The trusted browser harness is separate from the application. Stagehand explores and recovers browser actions; Playwright executes the semantic flow for the selected browser/viewport matrix. Each target receives its own recording. Chromium and Firefox with desktop and mobile viewports create four targets. A passed desktop flow does not fill missing mobile coverage. Provider retries are bounded by the run’s time, token, and cost limits and do not provide unlimited extra attempts.

​
4. Analyze and verify

Deterministic checks and image analysis inspect captured evidence. Candidate findings can be checked against fresh reproductions on the same target. Verified findings, inconclusive observations, and failed verification remain distinguishable. Brand findings require supplied rules or references. Optional checks add bounded diagnostics and, when enabled and eligible, video observations or paired base/head images. Missing required image analysis makes the review incomplete; a disabled optional check is unmeasured.

​
5. Deliver evidence and settle the run

Artifacts are stored privately and linked from the authenticated viewer. PR reviews update their GitHub Check and summary. Commit and local-source reviews have no PR Check or comment publication. Review execution, findings, and delivery status are separate: completing a browser flow does not prove publication succeeded. A completed, delivered review can consume credits even when it finds an aberration or produces a failing Check. Unsuccessful delivery releases the reservation under the billing settlement rules. See credits. The contributor local harness has a different scope: Codex planning, local execution and deterministic analysis, with local evidence and no GitHub publication or OpenRouter calls.