Optional checks

Enable checks

Open Repos → repository → Checks. Pro workspace administrators can enable performance, accessibility, contextual video review, base/head comparison, SEO, and link health. Performance and accessibility are on by default for eligible repositories without a saved policy. Existing opt-outs are preserved. Other checks remain off by default. Organization settings can restrict which checks repositories may use. Saving affects future reviews. Each run records its effective selection and policy version. These checks are report-only; they do not independently change the review score or establish that a PR introduced a defect.

Performance

Captures navigation/resource timings and bounded top-frame long-task and long-animation-frame observations where the browser supports them. Script attribution and expensive resources help explain slow recorded moments. The web-vitals library also captures available TTFB, FCP, LCP, CLS and INP observations and timing attribution during the recorded journey. These are automated lab observations. Unsupported APIs, dropped observations, and incomplete checkpoints remain explicit. Overlapping task/frame durations must not be added together. This is not field Core Web Vitals measurement, a complete CPU profile, or a React profiler. Sensitive flows using secret references skip source profiling. Lighthouse also performs one separate page-load audit per review attempt, capped at 25 seconds, using an eligible Chromium recording’s successful app-origin route. The Performance panel shows its lab score, page-load metrics, optimization findings, estimated savings and affected resources. It runs after recording in a fresh browser session; it does not reuse the journey’s login state. Secret-using flows, other browser engines and reproduction runs skip this extra audit. The report records viewport, engine and measurement conditions. Timeouts and partial results remain visible and do not fail the recorded journey.

Accessibility

Runs bounded axe checks on captured page states and reports violation rules and items requiring review, with corrective guidance, target selectors and checkpoints. Scans include visited pages, selected clicks and a reserved final or failed state, within a bounded scan budget. Timeouts and scanner failures remain visible. Inspect Coverage for collected results. A passed automated scan is not a conformance assessment; keyboard, assistive-technology, and unvisited-state coverage remain separate.

SEO and link health

SEO samples visited pages for title, description, canonical, and robots metadata. Provide explicit route expectations so preview noindex is not mistaken for a production indexing problem. Optional crawl-document reads inspect only the reviewed origin’s robots.txt and sitemap.xml within bounds. Link health samples visible anchors and current-document fragments. Only exact permitted app-origin paths receive HEAD probes. External links, redirects, credentials, query strings, and unsafe or state-changing destinations are not followed. A 404/410 indicates a missing destination; authentication failures, rate limits, and unsupported HEAD remain uncertain. This is not a site-wide crawler.

diffraction.json

{
  "version": 1,
  "review": {
    "siteHealth": {
      "environment": "preview",
      "productionOrigin": "https://example.com",
      "routes": [{ "path": "/", "indexing": "index", "title": "Example" }],
      "crawlDocuments": true,
      "safeLinkPaths": ["/help"]
    }
  }
}

environment accepts preview, production, or private. Repository expectations do not enable checks; save the corresponding switches in Checks. Secret-reference flows skip site-health collection.

Contextual video review

Adds bounded video analysis when usable, model-eligible video is present and provider/budget gates pass. Observations retain time intervals, rule references, and evidence links. A passed final assertion can coexist with a transient issue in the recording. Missing, skipped, or failed analysis stays visible. This feature does not guarantee detection of brief defects. Temporal observations remain outside verified findings and score penalties. Video analysis is distinct from simply playing a recorded video and from native Playwright video capture.

Base/head comparisons

Runs the same semantic flow against exact PR base/head revisions in separate environments. Comparable checkpoints show paired images, changed-pixel counts, and bounded crops. State mismatches, unstable captures, missing actions, or incompatible images prevent a truthful pixel comparison. Enable Base/head comparison and supply a synthetic-data contract:

diffraction.json

{
  "version": 1,
  "review": {
    "comparison": {
      "privacy": "dedicated-synthetic-data",
      "state": { "testId": "review-state", "value": "fixture-v1" },
      "fixedTime": "2026-09-07T00:00:00Z",
      "locale": "en-US",
      "timezoneId": "UTC",
      "masks": ["[data-private]"],
      "maximumTargets": 1
    }
  }
}

Both revisions must expose data-testid="review-state" with the exact text fixture-v1 at captured checkpoints. Supply the expected fixture marker in your application, not just the configuration. The target limit is one by default and at most four; optional checkpointStepIds selects up to eight checkpoints. Pixel comparison itself makes no AI calls. It still requires execution resources, and the surrounding review can use AI. Changed pixels are observations, not verified regressions. Cross-release new/resolved issue tracking and arbitrary authenticated-state comparison are not established by this feature.

Get started

Configure reviews

Understand results

Workspace

Reference

Enable checks

Performance

Accessibility

SEO and link health

Contextual video review

Base/head comparisons

​Enable checks

​Performance

​Accessibility

​SEO and link health

​Contextual video review

​Base/head comparisons

Enable checks

Performance

Accessibility

SEO and link health

Contextual video review

Base/head comparisons