Identify collapsed values and open them in a viewer (#533)

apps/yaak-client/components/core/Dropdown.tsx

@@ -233,23 +233,33 @@ export const Dropdown = forwardRef<DropdownRef, DropdownProps>(function Dropdown
 
 export interface ContextMenuProps {
   triggerPosition: { x: number; y: number } | null;
+  /**
+   * The box the menu belongs to, when the trigger has one.
+   *
+   * Placement aligns to the trigger's left or right edge depending on the room beside it, so a
+   * menu opened from an element wants its real rect. Without one the position is treated as a
+   * zero-width point, which is right for a menu opened at the cursor and wrong for one opened
+   * from a button.
+   */
+  triggerRect?: Pick<DOMRect, "top" | "bottom" | "left" | "right">;
   className?: string;
   items: DropdownProps["items"];
   onClose: () => void;
 }
 
 export const ContextMenu = forwardRef<DropdownRef, ContextMenuProps>(function ContextMenu(
-  { triggerPosition, className, items, onClose },
+  { triggerPosition, triggerRect, className, items, onClose },
   ref,
 ) {
   const triggerShape = useMemo(
-    () => ({
-      top: triggerPosition?.y ?? 0,
-      bottom: triggerPosition?.y ?? 0,
-      left: triggerPosition?.x ?? 0,
-      right: triggerPosition?.x ?? 0,
-    }),
-    [triggerPosition],
+    () =>
+      triggerRect ?? {
+        top: triggerPosition?.y ?? 0,
+        bottom: triggerPosition?.y ?? 0,
+        left: triggerPosition?.x ?? 0,
+        right: triggerPosition?.x ?? 0,
+      },
+    [triggerPosition, triggerRect],
   );
 
   if (triggerPosition == null) return null;
@@ -259,6 +269,9 @@ export const ContextMenu = forwardRef<DropdownRef, ContextMenuProps>(function Co
       isOpen={true} // Always open because we return null if not
       className={className}
       defaultSelectedIndex={null}
+      // A menu opened from an element points back at it. One opened at the cursor has nothing
+      // to point at, so it goes without.
+      showTriangle={triggerRect != null}
       ref={ref}
       items={items}
       onClose={onClose}

apps/yaak-client/components/core/Editor/Editor.css

@@ -121,16 +121,24 @@
         @apply cursor-default;
       }
 
-      /* An icon button any tag can carry, like the copy on a large value */
+      /* The chevron on a tag that opens a menu. The tag itself is the button, so this is
+         only the affordance that says so. */
       .tag-action {
-        @apply inline-flex items-center align-middle ml-1 cursor-pointer!;
-        @apply text-text-subtlest hover:text-text;
+        @apply inline-flex items-center align-middle ml-1;
+        @apply text-text-subtlest;
 
         svg {
           @apply w-[0.9em] h-[0.9em];
         }
       }
 
+      /* A preview of the value a tag stands for. Sized before the image loads, so arriving
+         never changes the line's height. */
+      .tag-thumbnail {
+        @apply inline-block align-middle mr-1 rounded-xs;
+        @apply w-[1.1em] h-[1.1em] object-cover bg-surface-highlight;
+      }
+
       .fn {
         @apply inline-block;
         .fn-inner {

apps/yaak-client/components/core/Editor/largeValues.test.ts

@@ -1,15 +1,15 @@
 import { EditorState } from "@codemirror/state";
 import { jsonc } from "@shopify/lang-jsonc";
 import { text } from "./text/extension";
-import { describe, expect, test, vi } from "vite-plus/test";
+import { describe, expect, test } from "vite-plus/test";
 import {
+  COLLAPSE_MEDIA_CHARS,
   COLLAPSE_TOKEN_CHARS,
-  largeValueField,
+  collapseDecorations,
   largeValues,
   MAX_VISIBLE_LINE_CHARS,
 } from "./largeValues";
-
-vi.mock("../../../lib/copy", () => ({ copyToClipboard: () => {} }));
+import type { SniffedValue } from "./sniffValue";
 
 const BIG = "A".repeat(1_000_000);
 
@@ -19,9 +19,19 @@ const jsonState = (doc: string) => EditorState.create({ doc, extensions: [jsonc(
 /** Without a grammar, so only the column rule applies */
 const plainState = (doc: string) => EditorState.create({ doc, extensions: largeValues });
 
+/**
+ * The decorations as if the whole document were on screen.
+ *
+ * In the editor the plugin passes the viewport instead, which is the same call with narrower
+ * ranges — the rules themselves don't know the difference.
+ */
+function decorationsFor(state: EditorState) {
+  return collapseDecorations(state, [{ from: 0, to: state.doc.length }]);
+}
+
 function collapsedRanges(state: EditorState) {
   const ranges: { from: number; to: number }[] = [];
-  const iter = state.field(largeValueField).decorations.iter();
+  const iter = decorationsFor(state).iter();
   while (iter.value != null) {
     ranges.push({ from: iter.from, to: iter.to });
     iter.next();
@@ -191,6 +201,67 @@ describe("undelimited tokens", () => {
   });
 });
 
+describe("sniffing the collapsed value", () => {
+  /** The widget standing in for each collapse, in document order */
+  function widgets(state: EditorState) {
+    const found: { valueFrom: number; sniffed: SniffedValue | null }[] = [];
+    const iter = decorationsFor(state).iter();
+    while (iter.value != null) {
+      found.push(iter.value.spec.widget);
+      iter.next();
+    }
+    return found;
+  }
+
+  /** A base64 value that starts with a PNG signature and runs on well past the limits */
+  function pngBase64(length = 1_000_000) {
+    const bytes = new Uint8Array(length);
+    bytes.set([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a]);
+    let binary = "";
+    for (const byte of bytes) binary += String.fromCharCode(byte);
+    return btoa(binary);
+  }
+
+  test("names the format of a base64 value inside JSON", () => {
+    const state = jsonState(`{"image":"${pngBase64()}"}`);
+    expect(widgets(state)[0]?.sniffed).toMatchObject({ mime: "image/png", label: "PNG" });
+  });
+
+  test("names the format of a data URI", () => {
+    const doc = `{"image":"data:image/jpeg;base64,${BIG}"}`;
+    expect(widgets(jsonState(doc))[0]?.sniffed).toMatchObject({ label: "JPEG" });
+  });
+
+  test("names nothing when the value is just text", () => {
+    expect(widgets(jsonState(`{"image":"${BIG}"}`))[0]?.sniffed).toBeNull();
+    expect(widgets(plainState(BIG))[0]?.sniffed).toBeNull();
+  });
+
+  test("reads a column cut from the start of its line, not from the cut", () => {
+    // A body that is nothing but one base64 blob: the tail is hidden, but the value it
+    // belongs to begins at the start of the line, which is where the signature is
+    const state = plainState(pngBase64());
+    const [widget] = widgets(state);
+
+    expect(widget?.valueFrom).toBe(0);
+    expect(widget?.sniffed).toMatchObject({ label: "PNG" });
+  });
+
+  test("reads a token collapse from the value itself", () => {
+    const doc = `{"image":"${pngBase64()}"}`;
+    const [widget] = widgets(jsonState(doc));
+
+    // Just past the opening quote, so the signature is the first thing it sees
+    expect(widget?.valueFrom).toBe(doc.indexOf('"', doc.indexOf("image") + 6) + 1);
+    expect(widget?.sniffed).toMatchObject({ label: "PNG" });
+  });
+
+  test("names nothing when a column cut lands mid-line after other text", () => {
+    const state = plainState(`some prefix text ${pngBase64()}`);
+    expect(widgets(state)[0]?.sniffed).toBeNull();
+  });
+});
+
 describe("recomputing", () => {
   test("updates when the document changes", () => {
     const state = plainState('{"image":"short"}');
@@ -202,3 +273,101 @@ describe("recomputing", () => {
     expect(collapsedRanges(next)).toHaveLength(1);
   });
 });
+
+describe("media collapsing, below the length rules", () => {
+  /** A real PNG signature, at a size that no length rule would touch */
+  function png(bytes: number) {
+    const b = new Uint8Array(bytes);
+    b.set([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a]);
+    let binary = "";
+    for (const byte of b) binary += String.fromCharCode(byte);
+    return btoa(binary);
+  }
+
+  /** Well under every length threshold, so only the media rule can collapse it */
+  const SMALL = png(3_000);
+
+  test("collapses a small image on a short line", () => {
+    expect(SMALL.length).toBeLessThan(MAX_VISIBLE_LINE_CHARS);
+    expect(SMALL.length).toBeLessThan(COLLAPSE_TOKEN_CHARS);
+
+    const state = jsonState(`{\n  "avatar": "${SMALL}"\n}`);
+    const ranges = collapsedRanges(state);
+
+    expect(ranges).toHaveLength(1);
+    expect(state.sliceDoc(ranges[0]!.from, ranges[0]!.to)).toBe(SMALL);
+  });
+
+  test("names it, so the tag can say what it is", () => {
+    const state = jsonState(`{"avatar":"${SMALL}"}`);
+    const iter = decorationsFor(state).iter();
+    expect(iter.value?.spec.widget.sniffed).toMatchObject({ mime: "image/png", label: "PNG" });
+  });
+
+  test("leaves a value of the same size alone when nothing recognises it", () => {
+    // The only difference from the case above is what the bytes turn out to be
+    const prose = "x".repeat(SMALL.length);
+    expect(collapsedRanges(jsonState(`{"note":"${prose}"}`))).toEqual([]);
+  });
+
+  test("leaves anything under the media threshold alone, recognised or not", () => {
+    const tiny = png(400);
+    expect(tiny.length).toBeLessThan(COLLAPSE_MEDIA_CHARS);
+    expect(collapsedRanges(jsonState(`{"icon":"${tiny}"}`))).toEqual([]);
+  });
+
+  test("leaves text alone even when its first bytes match a signature", () => {
+    // `Qk0` decodes to `BM`, the whole of the BMP signature. Two bytes come up by chance often
+    // enough that the sniff alone must not be allowed to hide something.
+    const prose = `Qk0 ${"the quick brown fox jumps over the lazy dog. ".repeat(40)}`;
+    expect(prose.length).toBeGreaterThan(COLLAPSE_MEDIA_CHARS);
+    expect(collapsedRanges(jsonState(`{"note":"${prose}"}`))).toEqual([]);
+  });
+
+  test("still collapses it once it is long enough to be a rendering problem", () => {
+    // Past the length rule the sniff no longer decides anything, so this is hidden for its size
+    const prose = "Qk0 " + "words and more words ".repeat(COLLAPSE_TOKEN_CHARS);
+    expect(collapsedRanges(jsonState(`{"note":"${prose}"}`))).toHaveLength(1);
+  });
+
+  test("collapses a data URI on a short line", () => {
+    const state = jsonState(`{"avatar":"data:image/jpeg;base64,${SMALL}"}`);
+    expect(collapsedRanges(state)).toHaveLength(1);
+  });
+
+  test("still leaves an ordinary document completely untouched", () => {
+    const doc = Array.from({ length: 5_000 }, (_, i) => `  { "id": ${i}, "name": "item" },`).join(
+      "\n",
+    );
+    expect(collapsedRanges(jsonState(doc))).toEqual([]);
+  });
+});
+
+describe("viewport scoping", () => {
+  const doc = () => {
+    const value = "A".repeat(COLLAPSE_TOKEN_CHARS * 2);
+    return `{\n${["a", "b", "c"].map((k) => `  "${k}": "${value}"`).join(",\n")}\n}`;
+  };
+
+  test("decorates only the ranges it is given", () => {
+    const state = jsonState(doc());
+    const secondLine = state.doc.line(3);
+
+    const all = collapseDecorations(state, [{ from: 0, to: state.doc.length }]);
+    const one = collapseDecorations(state, [{ from: secondLine.from, to: secondLine.to }]);
+
+    expect(all.size).toBe(3);
+    expect(one.size).toBe(1);
+  });
+
+  test("covers a line the range only partly overlaps", () => {
+    // The viewport can start mid-line; the collapse still has to span the whole value
+    const state = jsonState(doc());
+    const line = state.doc.line(2);
+    const partial = collapseDecorations(state, [{ from: line.from + 5, to: line.from + 6 }]);
+
+    expect(partial.size).toBe(1);
+    const iter = partial.iter();
+    expect(iter.to).toBeGreaterThan(line.from + 6);
+  });
+});

apps/yaak-client/components/core/Editor/largeValues.ts

@@ -1,11 +1,12 @@
 import { ensureSyntaxTree, syntaxTree } from "@codemirror/language";
 import { formatSize } from "@yaakapp-internal/lib/formatSize";
 import type { EditorState, Extension, Range } from "@codemirror/state";
-import { StateField } from "@codemirror/state";
 import type { Tree as SyntaxTree } from "@lezer/common";
-import type { DecorationSet } from "@codemirror/view";
-import { Decoration, EditorView, WidgetType } from "@codemirror/view";
-import { copyToClipboard } from "../../../lib/copy";
+import type { DecorationSet, ViewUpdate } from "@codemirror/view";
+import { Decoration, EditorView, ViewPlugin, WidgetType } from "@codemirror/view";
+import { fireAndForget } from "../../../lib/fireAndForget";
+import type { SniffedValue } from "./sniffValue";
+import { isEncodedRun, SNIFF_HEAD_CHARS, sniffValue } from "./sniffValue";
 
 /**
  * How much of a line may be rendered before the rest is collapsed.
@@ -17,53 +18,131 @@ import { copyToClipboard } from "../../../lib/copy";
 export const MAX_VISIBLE_LINE_CHARS = 10_000;
 
 /**
- * A quoted value longer than this on an over-long line is collapsed whole, ahead of the column
- * cut, so the structure around it stays visible. Needs a grammar to find.
+ * A quoted value longer than this is collapsed whole, whatever it turns out to hold. Long
+ * enough that nothing a person might actually read is ever hidden on length alone.
  */
 export const COLLAPSE_TOKEN_CHARS = 5_000;
 
 /**
- * Keeps over-long lines from reaching layout, which is what makes the editor stall on a
- * base64 blob or a minified payload.
+ * A quoted value longer than this is collapsed too, but only when {@link sniffValue} can say
+ * what it is. Encoded media is never worth reading, so once we can name it and offer a viewer,
+ * a tag beats a wall of base64 well below the length that would make it a rendering problem —
+ * this covers the avatars and icons that make up most base64 in real payloads. Anything we
+ * can't identify stays visible until it's big enough to be a problem on its own.
+ */
+export const COLLAPSE_MEDIA_CHARS = 1_000;
+
+/**
+ * Collapses what can't be read: over-long lines, which stall layout, and encoded media, which
+ * is only noise on screen.
  *
- * Cost tracks the length of the longest line, not the size of the document. A 1 MB response of
- * ordinary multi-line JSON renders fine, while the same 1 MB on a single line stalls the UI,
- * because soft wrap has to measure the whole line end to end to find its break points.
+ * Layout cost tracks the length of the longest line, not the size of the document. A 1 MB
+ * response of ordinary multi-line JSON renders fine, while the same 1 MB on a single line stalls
+ * the UI, because soft wrap has to measure the whole line end to end to find its break points.
  * Measured in WKWebView, the engine macOS ships: 221 ms per render pass at 1 MB and 968 ms at
  * 3 MB, against 51 ms and 138 ms once collapsed, with soft wrap left on.
  *
- * Two rules, applied only to lines over {@link MAX_VISIBLE_LINE_CHARS}, so ordinary documents
- * are untouched:
+ * Three rules:
  *
  * 1. Collapse quoted values over {@link COLLAPSE_TOKEN_CHARS} whole, so a base64 string reads
- *    as `"image": "999.8 KB hidden…"` with its key intact. A minified body with several large values
- *    keeps every one of its keys. Needs a grammar to find the value.
- * 2. Collapse whatever is still past the column limit. This needs no grammar, so it covers
- *    plain text, undelimited tokens, and any line that is simply long.
+ *    as `"image": "PNG · 999.8 KB"` with its key intact. A minified body with several large
+ *    values keeps every one of its keys. Needs a grammar to find the value.
+ * 2. Collapse quoted values over {@link COLLAPSE_MEDIA_CHARS} that sniff as a known format,
+ *    which is what catches an ordinary embedded image long before it is a rendering problem.
+ * 3. On a line still over {@link MAX_VISIBLE_LINE_CHARS}, collapse whatever is left past the
+ *    column limit. This needs no grammar, so it covers plain text, undelimited tokens, and any
+ *    line that is simply long.
+ *
+ * Only what the viewport covers is examined, and only lines long enough to hold a collapse are
+ * looked at, so a document of ordinary short lines costs a length check per visible line and
+ * nothing else — no grammar is consulted and no tree is walked.
  *
  * Nothing leaves the document. Copy, filter and save all still see the full text, and the tag
- * carries a button that copies exactly what it hides.
+ * itself is a button, opening a menu that views, copies or saves exactly what it stands for.
+ *
+ * The value's own head says what it is — a `data:` URI names its media type, and raw base64
+ * gives its format up in the first dozen bytes — so the tag can name the format and open the
+ * value in the matching viewer without any of it being read. See {@link sniffValue}.
  *
  * Read-only editors only. Hiding part of a document someone is editing would mean editing
  * text they can't see.
  */
 
-/** A hidden range. The tag stands in for exactly this text, and the dialog shows it. */
+/**
+ * A hidden range, the value it belongs to, and what that value turned out to be.
+ *
+ * The hidden range and the value differ under the column rule, where the value starts at the
+ * beginning of the line but only what runs past the limit is hidden. The tag stands in for
+ * [from, to) and copies exactly that; the viewer needs the whole value, [valueFrom, to).
+ */
 interface Collapse {
   from: number;
   to: number;
+  valueFrom: number;
+  sniffed: SniffedValue | null;
+}
+
+/**
+ * Lucide's `chevron-down`, inlined because the widget builds its DOM synchronously and
+ * rendering React here would leave it empty while CodeMirror measures line heights.
+ */
+const CHEVRON_DOWN =
+  '<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="none" ' +
+  'stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" ' +
+  'aria-hidden="true"><path d="m6 9 6 6 6-6"/></svg>';
+
+/**
+ * Marks the tag as the thing that opens the menu.
+ *
+ * The whole tag is the target rather than the chevron alone: there is only one action, and a
+ * label you can't click next to a hit area a few pixels wide is a worse button than the tag
+ * itself. The chevron stays as the affordance that says so.
+ *
+ * A span with a role rather than a real button, because a button's box model makes the line
+ * taller — the one thing this extension exists to keep from happening.
+ */
+function makeTagButton(el: HTMLElement, title: string, onOpen: (at: DOMRect) => void) {
+  el.role = "button";
+  el.ariaHasPopup = "menu";
+  el.tabIndex = 0;
+  el.title = title;
+  el.ariaLabel = title;
+
+  // Keep the editor from moving the cursor when the tag is pressed
+  el.addEventListener("mousedown", (e) => {
+    e.preventDefault();
+    e.stopPropagation();
+  });
+  el.addEventListener("click", (e) => {
+    e.preventDefault();
+    e.stopPropagation();
+    onOpen(el.getBoundingClientRect());
+  });
+  el.addEventListener("keydown", (e) => {
+    if (e.key !== "Enter" && e.key !== " ") return;
+    e.preventDefault();
+    e.stopPropagation();
+    onOpen(el.getBoundingClientRect());
+  });
 }
 
 class LargeValueWidget extends WidgetType {
   constructor(
     private readonly from: number,
     private readonly to: number,
+    private readonly valueFrom: number,
+    private readonly sniffed: SniffedValue | null,
   ) {
     super();
   }
 
   eq(other: LargeValueWidget) {
-    return other.from === this.from && other.to === this.to;
+    return (
+      other.from === this.from &&
+      other.to === this.to &&
+      other.valueFrom === this.valueFrom &&
+      other.sniffed?.mime === this.sniffed?.mime
+    );
   }
 
   toDOM(view: EditorView) {
@@ -75,68 +154,166 @@ class LargeValueWidget extends WidgetType {
     // tag-sized border is meant to be near invisible.
     el.className = "x-theme-templateTag x-theme-templateTag--secondary template-tag";
 
+    const thumbnail = buildThumbnail(view, this.valueFrom, this.to, this.sniffed);
+    if (thumbnail != null) {
+      el.appendChild(thumbnail);
+    }
+
     const label = document.createElement("span");
-    label.textContent = `${formatSize(length)} hidden…`;
+    // A named type reads as a stand-in for the value, so it only needs to say what and how big.
+    // Without a name there is nothing to show but the fact that something is missing.
+    label.textContent =
+      this.sniffed == null
+        ? `${formatSize(length)} hidden…`
+        : `${this.sniffed.label} · ${formatSize(length)}`;
     el.appendChild(label);
 
-    // A span rather than a button: a button's box model makes the line taller. It keeps the
-    // role and name so assistive tech still sees it as an action.
-    const copy = document.createElement("span");
-    copy.role = "button";
-    copy.className = "tag-action";
-    copy.title = `Copy hidden text (${length.toLocaleString()} characters)`;
-    copy.ariaLabel = copy.title;
-    // Lucide's `copy` icon, inlined because the widget builds its DOM synchronously and
-    // rendering React here would leave it empty while CodeMirror measures line heights
-    copy.innerHTML =
-      '<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="none" ' +
-      'stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" ' +
-      'aria-hidden="true"><rect width="14" height="14" x="8" y="8" rx="2" ry="2"/>' +
-      '<path d="M4 16c-1.1 0-2-.9-2-2V4c0-1.1.9-2 2-2h10c1.1 0 2 .9 2 2"/></svg>';
-    // Keep the editor from moving the cursor when the button is pressed
-    copy.addEventListener("mousedown", (e) => {
-      e.preventDefault();
-      e.stopPropagation();
-    });
-    copy.addEventListener("click", (e) => {
-      e.preventDefault();
-      e.stopPropagation();
-      copyToClipboard(view.state.sliceDoc(this.from, this.to));
-    });
-    el.appendChild(copy);
+    const chevron = document.createElement("span");
+    chevron.className = "tag-action";
+    chevron.ariaHidden = "true";
+    chevron.innerHTML = CHEVRON_DOWN;
+    el.appendChild(chevron);
+
+    makeTagButton(
+      el,
+      this.sniffed == null ? "Value actions" : `${this.sniffed.label} actions`,
+      (rect) => fireAndForget(this.openMenu(view, rect)),
+    );
 
     return el;
   }
 
+  /**
+   * Copy, view and save, in a menu opened against the button.
+   *
+   * Everything the menu needs is pulled in on click. The editor loads on every response, and
+   * neither the React menu nor the viewers behind the dialog are worth carrying until someone
+   * asks for them.
+   */
+  private async openMenu(
+    view: EditorView,
+    rect: Pick<DOMRect, "top" | "bottom" | "left" | "right">,
+  ) {
+    const [{ showContextMenu }, { showLargeValueDialog }, { encodingLabel, largeValueActions }] =
+      await Promise.all([
+        import("../../../lib/contextMenu"),
+        import("../../LargeValueDialog"),
+        import("../../../lib/largeValue"),
+      ]);
+
+    const { sniffed } = this;
+    // Sliced when an action runs rather than now, so opening the menu never touches the value
+    const value = () => view.state.sliceDoc(this.valueFrom, this.to);
+    const head = view.state.sliceDoc(
+      this.valueFrom,
+      Math.min(this.valueFrom + SNIFF_HEAD_CHARS, this.to),
+    );
+
+    showContextMenu({
+      id: "large-value",
+      // The whole tag, so the menu can align to whichever edge has room beside it
+      triggerPosition: { x: rect.left, y: rect.bottom },
+      triggerRect: rect,
+      items: [
+        { type: "separator", label: encodingLabel(head, sniffed, this.to - this.from) },
+        ...largeValueActions({
+          value,
+          sniffed,
+          // Exactly what the tag stands in for, which under the column rule is only the tail
+          copyText: () => view.state.sliceDoc(this.from, this.to),
+          onView: () => showLargeValueDialog({ text: value(), sniffed }),
+        }),
+      ],
+    });
+  }
+
   ignoreEvent() {
     return false;
   }
 }
 
+/**
+ * How much encoded image a tag will preview. Roughly 3 MB of file, base64 being 4 bytes to 3.
+ *
+ * The decode runs on WebKit's image thread, but the bitmap it produces is sized by the image's
+ * own dimensions, not by the box we draw it in — a 4000×3000 photo costs 48 MB however small the
+ * thumbnail. So there has to be a limit, even though encoded size is only a proxy for the one
+ * that matters: a well-compressed photo can decode larger than a lossless screenshot twice its
+ * file size. This is set high enough to cover ordinary screenshots and wallpapers, since a tag
+ * that previews some images and not others is worse than one that previews none. Only widgets in
+ * the viewport are ever built, so a screenful is the most that decode at once.
+ */
+const THUMBNAIL_MAX_CHARS = 4_000_000;
+
+/**
+ * A preview of the value, at a size fixed before it loads.
+ *
+ * The box is a fixed square from the moment it's inserted, so the image arriving never changes
+ * the line's height or the tag's width. A growing line box is the layout cost this whole
+ * extension exists to avoid, and an image of unknown dimensions is the classic way to cause one.
+ *
+ * The src is the value's own text handed straight to the decoder as a data URI. Decoding the
+ * base64 ourselves first would mean walking megabytes on the main thread, which is the one
+ * thing that must not happen here.
+ */
+function buildThumbnail(
+  view: EditorView,
+  valueFrom: number,
+  to: number,
+  sniffed: SniffedValue | null,
+): HTMLElement | null {
+  if (sniffed == null || !sniffed.mime.startsWith("image/")) return null;
+  if (sniffed.encoding !== "base64") return null;
+  if (to - valueFrom > THUMBNAIL_MAX_CHARS) return null;
+
+  const img = document.createElement("img");
+  img.className = "tag-thumbnail";
+  img.alt = "";
+  img.ariaHidden = "true";
+  img.decoding = "async";
+  // A magic number can be wrong, and a data URI can lie. Drop the box rather than leave a
+  // broken-image glyph sitting in the tag.
+  img.addEventListener("error", () => img.remove());
+  // Read the document and build the data URI after the widget is measured and on screen
+  requestAnimationFrame(() => {
+    if (!img.isConnected) return;
+    const payload = view.state.sliceDoc(valueFrom + sniffed.offset, to);
+    img.src = `data:${sniffed.mime};base64,${payload}`;
+  });
+  return img;
+}
+
 interface Line {
   from: number;
   to: number;
 }
 
-/** Lines long enough to be a problem. Most documents have none, and we stop there. */
-function findLongLines(text: string): Line[] {
-  if (text.length <= MAX_VISIBLE_LINE_CHARS) {
-    return []; // No line can be longer than the whole text
+/**
+ * The lines the given ranges touch, in document order and each listed once.
+ *
+ * Only lines long enough to hold a collapse are returned. That check is what keeps this free on
+ * ordinary documents: no grammar is consulted and no tree is walked for a screen of short lines.
+ */
+function collapsibleLines(state: EditorState, ranges: readonly { from: number; to: number }[]) {
+  const lines: Line[] = [];
+  let lastFrom = -1;
+
+  for (const range of ranges) {
+    let pos = range.from;
+    for (;;) {
+      const line = state.doc.lineAt(pos);
+      if (line.from > lastFrom) {
+        lastFrom = line.from;
+        if (line.to - line.from >= COLLAPSE_MEDIA_CHARS) {
+          lines.push({ from: line.from, to: line.to });
+        }
+      }
+      if (line.to >= range.to) break;
+      pos = line.to + 1;
+    }
   }
 
-  const lines: Line[] = [];
-  let from = 0;
-  for (;;) {
-    const newline = text.indexOf("\n", from);
-    const to = newline < 0 ? text.length : newline;
-    if (to - from > MAX_VISIBLE_LINE_CHARS) {
-      lines.push({ from, to });
-    }
-    if (newline < 0) {
-      return lines;
-    }
-    from = newline + 1;
-  }
+  return lines;
 }
 
 /**
@@ -148,33 +325,44 @@ function findLongLines(text: string): Line[] {
  */
 const PARSE_TIMEOUT_MS = 100;
 
-/** The parsed tree covering the long lines, as far as parsing got in the time allowed. */
-function treeForLongLines(state: EditorState, longLines: Line[]): SyntaxTree {
-  const lastLine = longLines[longLines.length - 1];
-  if (lastLine == null) {
+/**
+ * The parsed tree covering these lines.
+ *
+ * Parsing is only ever forced for a line long enough to stall layout, where finding the values
+ * inside it is what saves the frame. For everything else we take whatever the background parse
+ * has reached and decorate again when it advances — a forced parse runs from the top of the
+ * document, so making one happen on every scroll would cost far more than the tags are worth.
+ */
+function treeForLines(state: EditorState, lines: Line[]): SyntaxTree {
+  const last = lines[lines.length - 1];
+  const stalls = lines.some((l) => l.to - l.from > MAX_VISIBLE_LINE_CHARS);
+  if (last == null || !stalls) {
     return syntaxTree(state);
   }
-  return ensureSyntaxTree(state, lastLine.to, PARSE_TIMEOUT_MS) ?? syntaxTree(state);
+  return ensureSyntaxTree(state, last.to, PARSE_TIMEOUT_MS) ?? syntaxTree(state);
 }
 
 const QUOTES = ['"', "'", "`"];
 
 /**
- * Quoted values on this line big enough to collapse whole, in document order.
+ * Quoted values on this line worth collapsing, in document order.
  *
  * Everything between the quotes goes, so a long string reads as `"image": "<tag>"` and the key
  * still tells you what it is. Undelimited tokens are left to the column cut instead: replacing
  * one whole would leave the line with nothing on it but a tag.
  */
-function findLargeTokens(state: EditorState, tree: SyntaxTree, line: Line): Collapse[] {
+function findCollapsibleTokens(state: EditorState, tree: SyntaxTree, line: Line): Collapse[] {
   const collapses: Collapse[] = [];
+  // Length alone only hides things on a line that would stall without it. On a line that renders
+  // fine, being long is not a reason to hide something we can't even name.
+  const stalls = line.to - line.from > MAX_VISIBLE_LINE_CHARS;
 
   tree.iterate({
     from: line.from,
     to: line.to,
     enter: (node) => {
-      // A node this small can't contain anything worth collapsing
-      if (node.to - node.from < COLLAPSE_TOKEN_CHARS) return false;
+      // A node this small can't contain anything worth collapsing, and neither can its children
+      if (node.to - node.from < COLLAPSE_MEDIA_CHARS) return false;
       // Only leaves, so we collapse the string itself rather than the object holding it
       if (node.node.firstChild != null) return true;
 
@@ -182,9 +370,25 @@ function findLargeTokens(state: EditorState, tree: SyntaxTree, line: Line): Coll
       const to = Math.min(node.to, line.to);
       const open = state.sliceDoc(from, from + 1);
       const quoted = to - from >= 2 && QUOTES.includes(open) && state.sliceDoc(to - 1, to) === open;
+      if (!quoted) return false;
 
-      if (quoted && to - 1 > from + 1) {
-        collapses.push({ from: from + 1, to: to - 1 });
+      // Everything inside the quotes. The whole value is hidden, so it is its own value range
+      const valueFrom = from + 1;
+      const valueTo = to - 1;
+      if (valueTo <= valueFrom) return false;
+
+      // The head is all it takes, so this never reads the value it describes
+      const head = state.sliceDoc(valueFrom, Math.min(valueFrom + SNIFF_HEAD_CHARS, valueTo));
+      const sniffed = sniffValue(head);
+
+      // Big enough to be a rendering problem on its own, or something we can both name and be
+      // sure is encoded — a short magic number matches by chance, and hiding text would be worse
+      // than showing base64
+      const oversized = stalls && valueTo - valueFrom >= COLLAPSE_TOKEN_CHARS;
+      const media =
+        sniffed != null && sniffed.encoding === "base64" && isEncodedRun(head, sniffed.offset);
+      if (oversized || media) {
+        collapses.push({ from: valueFrom, to: valueTo, valueFrom, sniffed });
       }
       return false;
     },
@@ -218,7 +422,11 @@ function findColumnCut(line: Line, tokens: Collapse[]): number {
 }
 
 function collapsesForLine(state: EditorState, tree: SyntaxTree, line: Line): Collapse[] {
-  const tokens = findLargeTokens(state, tree, line);
+  const tokens = findCollapsibleTokens(state, tree, line);
+  if (line.to - line.from <= MAX_VISIBLE_LINE_CHARS) {
+    return tokens; // Short enough to render, so there is nothing left to cut
+  }
+
   const cut = findColumnCut(line, tokens);
   if (cut < 0) {
     return tokens;
@@ -226,56 +434,69 @@ function collapsesForLine(state: EditorState, tree: SyntaxTree, line: Line): Col
 
   // The cut always lands in a visible stretch, so it never splits a token collapse
   const kept = tokens.filter((t) => t.to <= cut);
-  kept.push({ from: cut, to: line.to });
+  // Only the tail is hidden, but the value it belongs to runs from the start of the line — a
+  // body that is nothing but one base64 blob is still recognisable from there
+  const head = state.sliceDoc(line.from, Math.min(line.from + SNIFF_HEAD_CHARS, line.to));
+  kept.push({ from: cut, to: line.to, valueFrom: line.from, sniffed: sniffValue(head) });
   return kept;
 }
 
-function buildDecorations(state: EditorState, longLines: Line[]): DecorationSet {
-  if (longLines.length === 0) {
+/**
+ * The collapses covering the given document ranges.
+ *
+ * Kept separate from the plugin so the rules can be exercised against a state directly, with no
+ * view and no DOM.
+ */
+export function collapseDecorations(
+  state: EditorState,
+  ranges: readonly { from: number; to: number }[],
+): DecorationSet {
+  const lines = collapsibleLines(state, ranges);
+  if (lines.length === 0) {
     return Decoration.none;
   }
 
-  const tree = treeForLongLines(state, longLines);
-  const ranges: Range<Decoration>[] = [];
-  for (const line of longLines) {
-    for (const { from, to } of collapsesForLine(state, tree, line)) {
-      ranges.push(Decoration.replace({ widget: new LargeValueWidget(from, to) }).range(from, to));
+  const tree = treeForLines(state, lines);
+  const decorations: Range<Decoration>[] = [];
+  for (const line of lines) {
+    for (const { from, to, valueFrom, sniffed } of collapsesForLine(state, tree, line)) {
+      const widget = new LargeValueWidget(from, to, valueFrom, sniffed);
+      decorations.push(Decoration.replace({ widget }).range(from, to));
     }
   }
-  return Decoration.set(ranges);
+  return Decoration.set(decorations);
 }
 
-interface LargeValueState {
-  longLines: Line[];
-  decorations: DecorationSet;
-}
+/**
+ * Scoped to the viewport, so the cost is a screenful however big the document is.
+ *
+ * Recomputed when the document changes, when the viewport moves, and when background parsing
+ * advances — a value can only be found once the grammar has reached it.
+ */
+const largeValuePlugin = ViewPlugin.fromClass(
+  class {
+    decorations: DecorationSet;
 
-export const largeValueField = StateField.define<LargeValueState>({
-  create(state) {
-    const longLines = findLongLines(state.doc.toString());
-    return { longLines, decorations: buildDecorations(state, longLines) };
-  },
-
-  update(value, tr) {
-    if (tr.docChanged) {
-      const longLines = findLongLines(tr.state.doc.toString());
-      return { longLines, decorations: buildDecorations(tr.state, longLines) };
+    constructor(view: EditorView) {
+      this.decorations = collapseDecorations(view.state, view.visibleRanges);
     }
-    // Parsing is incremental, so a long line may only become a known token later. Documents
-    // with no long line can never gain a collapse, so they skip this entirely.
-    if (value.longLines.length > 0 && syntaxTree(tr.startState) !== syntaxTree(tr.state)) {
-      return { ...value, decorations: buildDecorations(tr.state, value.longLines) };
-    }
-    return value;
-  },
 
-  provide: (f) => [
-    EditorView.decorations.from(f, (v) => v.decorations),
+    update(update: ViewUpdate) {
+      if (
+        update.docChanged ||
+        update.viewportChanged ||
+        syntaxTree(update.startState) !== syntaxTree(update.state)
+      ) {
+        this.decorations = collapseDecorations(update.view.state, update.view.visibleRanges);
+      }
+    }
+  },
+  {
+    decorations: (v) => v.decorations,
     // Step the cursor over a placeholder instead of stranding it inside
-    EditorView.atomicRanges.of(
-      (view) => view.state.field(f, false)?.decorations ?? Decoration.none,
-    ),
-  ],
-});
+    provide: (plugin) =>
+      EditorView.atomicRanges.of((view) => view.plugin(plugin)?.decorations ?? Decoration.none),
+  },
+);
 
-export const largeValues: Extension = [largeValueField];
+export const largeValues: Extension = [largeValuePlugin];

apps/yaak-client/components/core/Editor/sniffValue.test.ts

@@ -0,0 +1,170 @@
+import { describe, expect, test } from "vite-plus/test";
+import { decodeBase64Prefix, labelForMime, SNIFF_HEAD_CHARS, sniffValue } from "./sniffValue";
+
+/** A base64 value that starts with `magic` and runs on for a while, as a real one would */
+function base64Of(magic: number[], length = 2_000): string {
+  const bytes = new Uint8Array(length);
+  bytes.set(magic);
+  for (let i = magic.length; i < length; i++) {
+    bytes[i] = i % 251;
+  }
+  let binary = "";
+  for (const byte of bytes) {
+    binary += String.fromCharCode(byte);
+  }
+  return btoa(binary);
+}
+
+/** One character, one byte. These are all ASCII signatures. */
+function ascii(s: string): number[] {
+  const bytes: number[] = [];
+  for (let i = 0; i < s.length; i++) {
+    bytes.push(s.charCodeAt(i));
+  }
+  return bytes;
+}
+
+/** Four bytes no signature looks at: a RIFF chunk length, or an ISO-BMFF box size */
+const IGNORED = [0x00, 0x00, 0x00, 0x20];
+
+const PNG = [0x89, ...ascii("PNG"), 0x0d, 0x0a, 0x1a, 0x0a];
+const JPEG = [0xff, 0xd8, 0xff, 0xe0];
+const PDF = ascii("%PDF-1.7");
+
+describe("magic numbers", () => {
+  const cases: [string, number[], string, string][] = [
+    ["PNG", PNG, "image/png", "PNG"],
+    ["JPEG", JPEG, "image/jpeg", "JPEG"],
+    ["GIF", ascii("GIF89a"), "image/gif", "GIF"],
+    ["PDF", PDF, "application/pdf", "PDF"],
+    ["WEBP", [...ascii("RIFF"), ...IGNORED, ...ascii("WEBP")], "image/webp", "WEBP"],
+    ["WAV", [...ascii("RIFF"), ...IGNORED, ...ascii("WAVE")], "audio/wav", "WAV"],
+    ["ZIP", [...ascii("PK"), 0x03, 0x04], "application/zip", "ZIP"],
+    ["GZIP", [0x1f, 0x8b, 0x08], "application/gzip", "GZIP"],
+    ["MP3", [...ascii("ID3"), 0x04], "audio/mpeg", "MP3"],
+    ["OGG", ascii("OggS"), "audio/ogg", "OGG"],
+    ["FLAC", ascii("fLaC"), "audio/flac", "FLAC"],
+    ["WEBM", [0x1a, 0x45, 0xdf, 0xa3], "video/webm", "WEBM"],
+    ["MP4", [...IGNORED, ...ascii("ftypisom")], "video/mp4", "MP4"],
+    ["AVIF", [...IGNORED, ...ascii("ftypavif")], "image/avif", "AVIF"],
+    ["HEIC", [...IGNORED, ...ascii("ftypheic")], "image/heic", "HEIC"],
+    ["M4A", [...IGNORED, ...ascii("ftypM4A ")], "audio/mp4", "M4A"],
+    ["MOV", [...IGNORED, ...ascii("ftypqt  ")], "video/quicktime", "MOV"],
+    ["BMP", ascii("BMxx"), "image/bmp", "BMP"],
+    ["TIFF", [...ascii("II"), 0x2a, 0x00], "image/tiff", "TIFF"],
+  ];
+
+  for (const [name, magic, mime, label] of cases) {
+    test(`recognises ${name}`, () => {
+      expect(sniffValue(base64Of(magic))).toEqual({ mime, label, offset: 0, encoding: "base64" });
+    });
+  }
+
+  test("recognises nothing in arbitrary base64", () => {
+    expect(sniffValue(btoa("just some text that happens to be encoded"))).toBeNull();
+  });
+
+  test("recognises nothing in text that isn't base64", () => {
+    expect(sniffValue("the quick brown fox jumps over the lazy dog".repeat(10))).toBeNull();
+  });
+
+  test("recognises nothing in a JWT-shaped value", () => {
+    expect(sniffValue(`${btoa('{"alg":"HS256"}')}.${btoa('{"sub":"1"}')}.c2ln`)).toBeNull();
+  });
+
+  test("reads the url-safe alphabet", () => {
+    const urlSafe = base64Of(PNG).replaceAll("+", "-").replaceAll("/", "_");
+    expect(sniffValue(urlSafe)?.label).toBe("PNG");
+  });
+
+  test("ignores leading whitespace, and counts it in the offset", () => {
+    expect(sniffValue(`   ${base64Of(PNG)}`)).toMatchObject({ label: "PNG", offset: 3 });
+  });
+});
+
+describe("data URIs", () => {
+  test("takes the declared type over the bytes", () => {
+    const uri = `data:image/svg+xml;base64,${btoa("<svg/>")}`;
+    expect(sniffValue(uri)).toEqual({
+      mime: "image/svg+xml",
+      label: "SVG",
+      offset: "data:image/svg+xml;base64,".length,
+      encoding: "base64",
+    });
+  });
+
+  test("points past the header, so the payload can be decoded from there", () => {
+    const payload = base64Of(PNG);
+    const uri = `data:image/png;base64,${payload}`;
+    const sniffed = sniffValue(uri)!;
+    expect(uri.slice(sniffed.offset, sniffed.offset + 8)).toBe(payload.slice(0, 8));
+  });
+
+  test("falls back to the bytes when the declared type says nothing", () => {
+    const uri = `data:application/octet-stream;base64,${base64Of(PDF)}`;
+    expect(sniffValue(uri)).toMatchObject({ mime: "application/pdf", label: "PDF" });
+  });
+
+  test("handles a header with no type at all", () => {
+    expect(sniffValue(`data:;base64,${base64Of(JPEG)}`)).toMatchObject({ mime: "image/jpeg" });
+  });
+
+  test("handles extra parameters", () => {
+    const uri = `data:text/plain;charset=utf-8;base64,${btoa("hello")}`;
+    expect(sniffValue(uri)).toMatchObject({ mime: "text/plain", encoding: "base64" });
+  });
+
+  test("handles a percent-encoded payload", () => {
+    expect(sniffValue("data:text/html,%3Ch1%3Ehi%3C/h1%3E")).toEqual({
+      mime: "text/html",
+      label: "HTML",
+      offset: "data:text/html,".length,
+      encoding: "percent",
+    });
+  });
+
+  test("defaults a bare percent-encoded payload to text", () => {
+    expect(sniffValue("data:,hello%20there")).toMatchObject({
+      mime: "text/plain",
+      encoding: "percent",
+    });
+  });
+});
+
+describe("decoding only the head", () => {
+  test("reads no more of the value than the bytes asked for", () => {
+    const big = base64Of(PNG, 3_000_000);
+    // Truncating to the head must not change what is found, which is the whole point:
+    // callers only ever hand it a slice
+    expect(sniffValue(big.slice(0, SNIFF_HEAD_CHARS))).toEqual(sniffValue(big));
+  });
+
+  test("decodes whole 4-character groups only", () => {
+    // 12 bytes needs 16 characters; a 15-character slice yields the 12 bytes of 3 whole groups
+    expect(decodeBase64Prefix(base64Of(PNG), 0, 12)).toHaveLength(12);
+    expect(decodeBase64Prefix(base64Of(PNG).slice(0, 15), 0, 12)).toHaveLength(9);
+  });
+
+  test("declines a value too short to hold a group", () => {
+    expect(decodeBase64Prefix("abc", 0, 12)).toBeNull();
+  });
+
+  test("declines text outside the alphabet", () => {
+    expect(decodeBase64Prefix("hello world, not base64!", 0, 12)).toBeNull();
+  });
+});
+
+describe("labels", () => {
+  test.each([
+    ["image/png", "PNG"],
+    ["image/svg+xml", "SVG"],
+    ["text/plain", "TEXT"],
+    ["application/vnd.ms-excel", "MS-EXCEL"],
+    ["image/x-icon", "ICON"],
+    ["application/octet-stream", "APPLICATION"],
+    ["font/woff2", "WOFF2"],
+    ["application/vnd.openxmlformats-officedocument.wordprocessingml.document", "DOCUMENT"],
+  ])("%s reads as %s", (mime, label) => {
+    expect(labelForMime(mime)).toBe(label);
+  });
+});

apps/yaak-client/components/core/Editor/sniffValue.ts

@@ -0,0 +1,219 @@
+/**
+ * What a collapsed value turns out to be.
+ *
+ * Sniffing reads the head of the value only. A `data:` URI states its own media type, and raw
+ * base64 gives up its type in the first few bytes, so neither needs the rest — which is the
+ * point, since the whole reason these values are collapsed is that touching all of one costs
+ * enough to stall the UI.
+ */
+export interface SniffedValue {
+  /** Full media type, e.g. `image/png`. What the dialog routes on. */
+  mime: string;
+  /** A word for the tag, e.g. `PNG`. */
+  label: string;
+  /** Where the encoded payload starts in the value, past any `data:` header. */
+  offset: number;
+  encoding: "base64" | "percent";
+}
+
+/**
+ * How much of a value {@link sniffValue} needs. Enough for a `data:` header of realistic
+ * length, and far more than the 24 base64 characters the magic numbers are read from.
+ */
+export const SNIFF_HEAD_CHARS = 256;
+
+/** Bytes the longest signature needs: `RIFF????WEBP`, and an ISO-BMFF brand at offset 8. */
+const SNIFF_BYTES = 12;
+
+/** `data:[<mime>][;<param>…],` — the payload follows the comma. */
+const DATA_URI = /^data:([^,;]*)((?:;[^,;]*)*),/;
+
+/** Media types that name no format, so the bytes are worth a look even when one is declared. */
+const UNINFORMATIVE = ["", "application/octet-stream", "binary/octet-stream"];
+
+interface Signature {
+  label: string;
+  mime: string;
+  offset?: number;
+  /** Byte values, where `null` matches anything */
+  magic: readonly (number | null)[];
+}
+
+/**
+ * A magic number written as the ASCII it reads as, with `?` for a byte that varies.
+ *
+ * Indexed rather than iterated, because a signature is a sequence of bytes: one character here
+ * is one byte, and splitting into code points would be the wrong unit for that.
+ */
+function ascii(pattern: string): (number | null)[] {
+  const bytes: (number | null)[] = [];
+  for (let i = 0; i < pattern.length; i++) {
+    bytes.push(pattern[i] === "?" ? null : pattern.charCodeAt(i));
+  }
+  return bytes;
+}
+
+const SIGNATURES: readonly Signature[] = [
+  { label: "PNG", mime: "image/png", magic: [0x89, ...ascii("PNG"), 0x0d, 0x0a, 0x1a, 0x0a] },
+  { label: "JPEG", mime: "image/jpeg", magic: [0xff, 0xd8, 0xff] },
+  { label: "GIF", mime: "image/gif", magic: ascii("GIF8") },
+  { label: "BMP", mime: "image/bmp", magic: ascii("BM") },
+  { label: "WEBP", mime: "image/webp", magic: ascii("RIFF????WEBP") },
+  { label: "TIFF", mime: "image/tiff", magic: [...ascii("II"), 0x2a, 0x00] },
+  { label: "TIFF", mime: "image/tiff", magic: [...ascii("MM"), 0x00, 0x2a] },
+  { label: "ICO", mime: "image/x-icon", magic: [0x00, 0x00, 0x01, 0x00] },
+  { label: "PDF", mime: "application/pdf", magic: ascii("%PDF-") },
+  { label: "WAV", mime: "audio/wav", magic: ascii("RIFF????WAVE") },
+  { label: "AVI", mime: "video/x-msvideo", magic: ascii("RIFF????AVI ") },
+  { label: "MP3", mime: "audio/mpeg", magic: ascii("ID3") },
+  // A bare MPEG frame header: 11 sync bits, then a layer III / II / I version pair
+  { label: "MP3", mime: "audio/mpeg", magic: [0xff, 0xfb] },
+  { label: "MP3", mime: "audio/mpeg", magic: [0xff, 0xf3] },
+  { label: "MP3", mime: "audio/mpeg", magic: [0xff, 0xf2] },
+  { label: "OGG", mime: "audio/ogg", magic: ascii("OggS") },
+  { label: "FLAC", mime: "audio/flac", magic: ascii("fLaC") },
+  { label: "WEBM", mime: "video/webm", magic: [0x1a, 0x45, 0xdf, 0xa3] },
+  { label: "ZIP", mime: "application/zip", magic: [...ascii("PK"), 0x03, 0x04] },
+  { label: "GZIP", mime: "application/gzip", magic: [0x1f, 0x8b] },
+  { label: "7Z", mime: "application/x-7z-compressed", magic: [...ascii("7z"), 0xbc, 0xaf, 0x27] },
+  { label: "RAR", mime: "application/vnd.rar", magic: ascii("Rar!") },
+  { label: "GLTF", mime: "model/gltf-binary", magic: ascii("glTF") },
+];
+
+/**
+ * ISO base media files all start `????ftyp`, so the format is in the brand that follows rather
+ * than in the signature itself. Anything unlisted is some flavour of MP4.
+ */
+const ISO_BRANDS: Record<string, { mime: string; label: string }> = {
+  avif: { mime: "image/avif", label: "AVIF" },
+  avis: { mime: "image/avif", label: "AVIF" },
+  heic: { mime: "image/heic", label: "HEIC" },
+  heix: { mime: "image/heic", label: "HEIC" },
+  hevc: { mime: "image/heic", label: "HEIC" },
+  mif1: { mime: "image/heif", label: "HEIF" },
+  msf1: { mime: "image/heif", label: "HEIF" },
+  "M4A ": { mime: "audio/mp4", label: "M4A" },
+  "qt  ": { mime: "video/quicktime", label: "MOV" },
+};
+
+function matches(bytes: Uint8Array, { magic, offset = 0 }: Signature): boolean {
+  if (bytes.length < offset + magic.length) return false;
+  return magic.every((b, i) => b == null || bytes[offset + i] === b);
+}
+
+function readAscii(bytes: Uint8Array, from: number, length: number): string {
+  return String.fromCharCode(...bytes.subarray(from, from + length));
+}
+
+/** The format the first bytes of a file identify it as, or null if they identify nothing. */
+export function sniffBytes(bytes: Uint8Array): { mime: string; label: string } | null {
+  if (readAscii(bytes, 4, 4) === "ftyp") {
+    return ISO_BRANDS[readAscii(bytes, 8, 4)] ?? { mime: "video/mp4", label: "MP4" };
+  }
+  const sig = SIGNATURES.find((s) => matches(bytes, s));
+  return sig == null ? null : { mime: sig.mime, label: sig.label };
+}
+
+/**
+ * The first `bytes` bytes of a base64 payload, or null if it isn't base64 after all.
+ *
+ * Only the characters those bytes need are decoded. The trailing partial group is dropped
+ * rather than padded, since `atob` rejects a group of one or two characters outright and we
+ * have no use for the byte a three-character group would add.
+ */
+export function decodeBase64Prefix(text: string, offset: number, bytes: number): Uint8Array | null {
+  const wanted = Math.ceil(bytes / 3) * 4;
+  let b64 = text.slice(offset, offset + wanted);
+  b64 = b64.slice(0, b64.length - (b64.length % 4));
+  if (b64.length === 0) return null;
+
+  // The URL-safe alphabet stands for the same bytes
+  b64 = b64.replaceAll("-", "+").replaceAll("_", "/");
+  if (!/^[A-Za-z0-9+/]+$/.test(b64)) return null;
+
+  try {
+    const binary = atob(b64);
+    const out = new Uint8Array(binary.length);
+    for (let i = 0; i < binary.length; i++) {
+      out[i] = binary.charCodeAt(i);
+    }
+    return out;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * What a value is, from its head alone, or null if nothing recognises it.
+ *
+ * `head` need only be the first {@link SNIFF_HEAD_CHARS} characters — pass more and the rest is
+ * ignored. The offsets in the result are relative to the start of the whole value.
+ */
+export function sniffValue(head: string): SniffedValue | null {
+  const lead = head.length - head.trimStart().length;
+  const value = head.slice(lead);
+
+  const dataUri = DATA_URI.exec(value);
+  if (dataUri != null) {
+    const declared = dataUri[1]!.toLowerCase();
+    const offset = lead + dataUri[0].length;
+    if (!dataUri[2]!.split(";").includes("base64")) {
+      // A percent-encoded payload states its own type or is text by definition
+      const mime = declared === "" ? "text/plain" : declared;
+      return { mime, label: labelForMime(mime), offset, encoding: "percent" };
+    }
+    if (!UNINFORMATIVE.includes(declared)) {
+      return { mime: declared, label: labelForMime(declared), offset, encoding: "base64" };
+    }
+    return sniffBase64At(head, offset);
+  }
+
+  return sniffBase64At(head, lead);
+}
+
+function sniffBase64At(head: string, offset: number): SniffedValue | null {
+  const bytes = decodeBase64Prefix(head, offset, SNIFF_BYTES);
+  const sniffed = bytes == null ? null : sniffBytes(bytes);
+  return sniffed == null ? null : { ...sniffed, offset, encoding: "base64" };
+}
+
+/** The shortest run of base64 that {@link isEncodedRun} will accept as convincing */
+const ENCODED_RUN_CHARS = 64;
+
+/**
+ * Whether a value is one unbroken run of base64, rather than text that merely opens like it.
+ *
+ * A magic number is only a few bytes, so short ones match by chance — `BM` is two, which comes
+ * up about once in every 65,000 values. That was harmless while a value had to be long enough
+ * to hurt rendering before anything was hidden, and the sniff only chose the label. It is not
+ * harmless when the sniff itself decides to hide something, so that rule asks for this as well:
+ * prose leaves the alphabet within a few characters, and an encoded blob never does.
+ */
+export function isEncodedRun(head: string, offset: number): boolean {
+  const run = head.slice(offset);
+  return run.length >= ENCODED_RUN_CHARS && /^[A-Za-z0-9+/\-_]+={0,2}$/.test(run);
+}
+
+/** Subtypes whose own name makes a poor label */
+const LABEL_OVERRIDES: Record<string, string> = {
+  "text/plain": "TEXT",
+};
+
+/**
+ * A word short enough for a tag. The distinguishing part of a subtype is its last dotted
+ * segment before any `+suffix`, so `image/svg+xml` reads as SVG and `application/vnd.ms-excel`
+ * as MS-EXCEL. Anything still too long falls back to the top-level type.
+ */
+export function labelForMime(mime: string): string {
+  const override = LABEL_OVERRIDES[mime];
+  if (override != null) {
+    return override;
+  }
+
+  const [type, subtype] = mime.split("/");
+  const word = (subtype ?? "").split("+")[0]!.split(".").pop()!.replace(/^x-/, "");
+  if (word.length > 0 && word.length <= 10 && word !== "octet-stream") {
+    return word.toUpperCase();
+  }
+  return (type ?? mime).toUpperCase();
+}