AI Summaries

Most visual diff tools give you a red-pixel overlay and leave you to figure out what changed. RegressionBot compares the two pages' DOM trees, groups the changed elements into labeled regions (A, B, C…), and writes one line per region that quotes what changed — so you can read a six-line summary instead of hunting through screenshots.

How it Works

After capturing both pages, RegressionBot runs three steps to produce a summary. The screenshots are kept as pictures for you to look at. Nothing compares their pixels.

Example Summary

The regressionbotSummary field is a structured JSON array — one element per labeled region, each with a label (the region letter) and a text line. The first item has an empty label: it is the one-line overview. Below is an example for the six-region job shown above.

In most cases, reading the summary is enough to approve or escalate a job. To see a change on the page, open its diffUrl. It goes to the inspect view, which outlines every change on both versions.

Content Pushed Down

Adding or removing content moves everything below it. A new block added mid-page pushes the rest of the page down. Pixel-diff tools can't tell this shift from a real change.

RegressionBot matches elements by what they are, not by where they sit. An element that was only pushed down by a change above it is the same element with the same text and styles, so it is not a change.

Automated AI Verdicts

A regression is not always a bug. It is often exactly what the code change set out to do. Pass the intent behind the change as runContext — the PR title and description, or the commit message — and every changed page is judged against it. Pass no intent and nothing is judged.

if (summary.intentAssessment.decision !== 'pass') process.exit(1);

The field-by-field reference — basis, coveredBy, how to write an intent, and how to approve only the pages the verdict cleared — lives in Intent and the Verdict.

Parsing regressionbotSummary: it is an array of { label, text }, where label is the region letter shown in the inspect view. The one-line overview, when present, comes first with an empty label, and a whole-page change is described by a single item with an empty label, so code that keys on the label must tolerate one. Results from before September 28, 2026 were described from screenshots and may carry two fixed strings — one for a region with no captured text, one for an image whose source and dimensions are unchanged — both meaning "look at this yourself". Jobs created before v2.1 carry a plain string in this field instead of an array.

Using Summaries

The regressionbotSummary field is available wherever job results are returned.

import { RegressionBot } from '@regressionbot/sdk';

const client = new RegressionBot(); // reads REGRESSIONBOT_API_KEY from env
const job = await client.runProject('my-project');

await job.waitForCompletion(5000, undefined, { waitForSummaries: true }); // waits for both job and AI summary
const summary = await job.getSummary();
for (const r of summary.regressions) {
  // Typecast during transition period to avoid compilation errors on legacy SDK type declarations
  const aiSummary = (r.regressionbotSummary as unknown) as Array<{ label: string; text: string }> | undefined;
  for (const item of aiSummary ?? []) {
    // item.label = "A" | "B" | "" (empty for wholesale changes)
    // item.text  = single-sentence description
    console.log(item.label ? `${item.label}: ${item.text}` : item.text);
  }
}

Short Screenshot URLs

To make reviewing visual summaries as frictionless as possible, RegressionBot returns shortened, clickable redirect links for all screenshot fields (such as baselineUrl, currentUrl, and diffUrl).

Instead of cluttering your terminal logs or CI pipeline stdout with 600-character pre-signed URLs, you receive clean, 50-character clickable links:

Links are valid for 1 hour after generation — long enough to paste into a PR comment, Slack message, or CI log and open later. Clicking the link generates a fresh pre-signed image URL on the fly, so there is no separate expiry to track.