feat(cli): snapshot --zoom and per-finding crops on check --snapshots

snapshot --zoom <selector|x,y,w,h> + --zoom-scale (default 3) crops via
Puppeteer clip at raised deviceScaleFactor — density changes, layout
never does. Selector resolves per frame with 24px padding; no match is
a loud error, and a frame whose clamped region is a sliver (element
collapsed or animated off-canvas) is skipped with a stderr note rather
than written as a useless few-pixel image.

check --snapshots additionally writes finding-NN-<code>.png crops for
error findings with bboxes (cap 12, deterministic re-seek in a second
session) and draws labeled annotation boxes on overview frames via a
transient overlay injected only after audits complete. Skill reference
gains the zoom workflow: check reports a finding, zoom into it, fix,
re-check.
This commit is contained in:
Miguel Angel Simon Sierra
2026-07-10 13:27:52 -04:00
parent 58f45ef758
commit f4cef54b8b
12 changed files with 856 additions and 31 deletions
+150 -6
View File
@@ -1,6 +1,10 @@
import { mkdirSync, writeFileSync } from "node:fs";
import { join } from "node:path";
import type { Page } from "puppeteer-core";
import {
captureRegionCrop,
openSettledCompositionPage,
padCropRegion,
resolveCliChromeGpuMode,
seekCompositionTimeline,
waitForPreferredSeekTarget,
@@ -14,10 +18,12 @@ import { serveStaticProjectHtml } from "./staticProjectServer.js";
import type {
AnchoredLayoutIssue,
CheckAnchor,
CheckAnnotationBox,
CheckAuditDriver,
CheckBbox,
CheckBrowserResult,
CheckFinding,
CheckFindingCropRequest,
CheckGeometryCandidate,
CheckOptions,
CheckSeverity,
@@ -36,6 +42,11 @@ const SEEK_OPTIONS = {
settleMs: 120,
};
// --zoom's default padding/scale, reused here so finding crops carry the same
// bit of surrounding context an agent gets from `snapshot --zoom`.
const FINDING_CROP_PADDING_PX = 24;
const FINDING_CROP_SCALE = 3;
interface RuntimeDraft {
code: string;
severity: CheckSeverity;
@@ -121,6 +132,53 @@ export async function runBrowserCheck(
}
}
/**
* `check --snapshots`'s per-finding evidence crops. Opens its own session
* (the main grid session already closed by the time findings are shaped) and
* re-seeks to each finding's sample time — renders are deterministic, so a
* fresh page at the same time reproduces the same pixels the grid audited.
*/
export async function captureFindingCrops(
project: ProjectDir,
options: CheckOptions,
requests: CheckFindingCropRequest[],
): Promise<string[]> {
if (requests.length === 0) return [];
const { bundleToSingleHtml } = await import("@hyperframes/core/compiler");
const html = await bundleToSingleHtml(project.dir);
const server = await serveStaticProjectHtml(project.dir, html, "Failed to bind check server");
let chromeBrowser: import("puppeteer-core").Browser | undefined;
const written: string[] = [];
try {
const session = await openSettledCompositionPage(html, server.url, {
renderReadyTimeoutMs: options.timeout,
renderReadyWarningSuffix: "capturing finding crops",
browserGpuMode: resolveCliChromeGpuMode(),
});
chromeBrowser = session.browser;
const page = session.page;
await waitForPreferredSeekTarget(page, 500);
const snapshotDir = join(project.dir, "snapshots");
mkdirSync(snapshotDir, { recursive: true });
for (const request of requests) {
await seekCompositionTimeline(page, request.time, SEEK_OPTIONS);
const canvas = await page.evaluate(() => ({
width: window.innerWidth,
height: window.innerHeight,
}));
const region = padCropRegion(request.bbox, canvas, FINDING_CROP_PADDING_PX);
const buffer = await captureRegionCrop(page, region, FINDING_CROP_SCALE);
writeFileSync(join(snapshotDir, request.filename), buffer);
written.push(join("snapshots", request.filename));
}
return written;
} finally {
await chromeBrowser?.close().catch(() => undefined);
await server.close();
}
}
function wireRuntimeListeners(page: Page, drafts: RuntimeDraft[], currentTime: () => number): void {
page.on("console", (message) => {
const type = message.type();
@@ -202,7 +260,7 @@ function createPageDriver(page: Page, setTime: (time: number) => void): CheckAud
collectMotionFrame: (time, selectors, scopes) =>
collectMotionFrame(page, time, selectors, scopes),
anchorMotionIssues: (issues) => anchorLayoutIssues(page, issues),
collectContrast: (time) => collectContrast(page, time),
collectContrast: (time, annotations) => collectContrast(page, time, annotations),
};
}
@@ -430,20 +488,36 @@ async function compositionBbox(page: Page): Promise<CheckBbox> {
});
}
async function collectContrast(page: Page, time: number): Promise<ContrastCapture> {
async function collectContrast(
page: Page,
time: number,
layoutAnnotations: CheckAnnotationBox[] = [],
): Promise<ContrastCapture> {
let prepared: PreparedContrast[] = [];
try {
prepared = parsePreparedContrast(await prepareContrast(page, time));
const screenshot = await page.screenshot({ encoding: "base64", type: "png" });
if (typeof screenshot !== "string") throw new Error("Contrast screenshot was not base64");
// This screenshot is the one contrast math is sampled from below — it must
// stay untouched by the annotation overlay (finishContrast reads real
// painted pixels), so annotation only ever happens on a SECOND shot.
const measurementShot = await page.screenshot({ encoding: "base64", type: "png" });
if (typeof measurementShot !== "string") throw new Error("Contrast screenshot was not base64");
const raw = await finishContrast(
page,
screenshot,
measurementShot,
time,
prepared.map((entry) => entry.raw),
);
const finished = raw.flatMap(parseFinishedContrast);
return { entries: joinContrastEntries(finished, prepared), pngBase64: screenshot };
const entries = joinContrastEntries(finished, prepared);
// Contrast failures are only known once measurement above completes, so
// they're appended to the layout-derived annotations passed in by the
// pipeline rather than being requested up front.
const annotations = [
...layoutAnnotations,
...contrastFailureAnnotations(entries, layoutAnnotations.length),
];
const pngBase64 = await captureOverviewShot(page, annotations, measurementShot);
return { entries, pngBase64 };
} finally {
await page
.evaluate(() => {
@@ -454,6 +528,76 @@ async function collectContrast(page: Page, time: number): Promise<ContrastCaptur
}
}
function contrastFailureAnnotations(
entries: ContrastAuditEntry[],
labelOffset: number,
): CheckAnnotationBox[] {
return entries
.filter((entry) => !entry.wcagAA)
.map((entry, index) => ({
label: `${labelOffset + index + 1} contrast_aa_failure`,
bbox: entry.bbox,
}));
}
const ANNOTATION_OVERLAY_ID = "__hyperframesCheckAnnotations";
/**
* `check --snapshots`'s overview-frame annotation: every audit for this
* sample time has already run (layout/geometry findings arrive via
* `annotations`; contrast failures were just measured above), so it's safe
* to draw labeled boxes and take one more shot without perturbing anything
* audits read. No-op (returns the plain screenshot) when there's nothing to
* annotate — the common case stays exactly as before this feature existed.
*/
export async function captureOverviewShot(
page: Page,
annotations: CheckAnnotationBox[],
fallbackScreenshot: string,
): Promise<string> {
if (annotations.length === 0) return fallbackScreenshot;
await injectAnnotationOverlay(page, annotations);
try {
const shot = await page.screenshot({ encoding: "base64", type: "png" });
return typeof shot === "string" ? shot : fallbackScreenshot;
} finally {
await removeAnnotationOverlay(page);
}
}
/** One small, self-contained DOM overlay: fixed-position labeled boxes over
* each finding's bbox. Injected immediately before the annotated overview
* shot and torn down immediately after (see `captureOverviewShot`), so it
* never leaks into any audit's DOM reads. */
async function injectAnnotationOverlay(page: Page, boxes: CheckAnnotationBox[]): Promise<void> {
await page.evaluate(
(items: CheckAnnotationBox[], overlayId: string) => {
const root = document.createElement("div");
root.id = overlayId;
root.style.cssText = "position:fixed;inset:0;z-index:2147483647;pointer-events:none;";
for (const item of items) {
const box = document.createElement("div");
box.style.cssText = `position:fixed;left:${item.bbox.x}px;top:${item.bbox.y}px;width:${item.bbox.width}px;height:${item.bbox.height}px;border:2px solid #ff2d55;box-sizing:border-box;`;
const label = document.createElement("div");
label.textContent = item.label;
label.style.cssText =
"position:absolute;top:-18px;left:0;background:#ff2d55;color:#fff;font:11px/16px monospace;padding:0 4px;white-space:nowrap;";
box.appendChild(label);
root.appendChild(box);
}
document.body.appendChild(root);
},
boxes,
ANNOTATION_OVERLAY_ID,
);
}
async function removeAnnotationOverlay(page: Page): Promise<void> {
await page.evaluate((overlayId: string) => {
document.getElementById(overlayId)?.remove();
}, ANNOTATION_OVERLAY_ID);
}
async function prepareContrast(page: Page, time: number): Promise<unknown[]> {
// Candidate-to-element provenance must be captured while the prepare restore list is live.
// fallow-ignore-next-line complexity