heygen-com · vanceingalls · Jun 8, 2026 · Jun 8, 2026 · Jun 8, 2026 · Jun 8, 2026
diff --git a/packages/core/src/parsers/hfIds.test.ts b/packages/core/src/parsers/hfIds.test.ts
@@ -0,0 +1,117 @@
+import { describe, it, expect } from "vitest";
+import { ensureHfIds, mintHfId } from "./hfIds.js";
+import { parseHTML } from "linkedom";
+
+function ids(html: string): string[] {
+  const { document } = parseHTML(html);
+  return Array.from(document.querySelectorAll("[data-hf-id]")).map(
+    (e) => e.getAttribute("data-hf-id") as string,
+  );
+}
+
+// data-hf-id of the first element matching `selector`.
+function idOf(html: string, selector: string): string | null {
+  const { document } = parseHTML(html);
+  return document.querySelector(selector)?.getAttribute("data-hf-id") ?? null;
+}
+
+const doc = (body: string) => `<!doctype html><html><body>${body}</body></html>`;
+
+describe("ensureHfIds", () => {
+  it("mints a hf- id on every editable element node in body", () => {
+    const html = `<!doctype html><html><body>
+      <div class="card"><h1>Hi</h1><img src="a.png"><span>x</span></div>
+    </body></html>`;
+    const out = ensureHfIds(html);
+    for (const id of ids(out)) expect(id).toMatch(/^hf-[a-z0-9]{4}$/);
+    // div, h1, img, span = 4 ids
+    expect(ids(out)).toHaveLength(4);
+  });
+
+  it("skips script/style/template/meta and head", () => {
+    const html = `<!doctype html><html><head><meta charset="utf-8"></head>
+      <body><script>1</script><style>.a{}</style><p>keep</p></body></html>`;
+    const out = ensureHfIds(html);
+    // only the <p> gets an id
+    expect(ids(out)).toHaveLength(1);
+    expect(out).not.toContain("<script data-hf-id");
+    expect(out).not.toContain("<style data-hf-id");
+    expect(out).not.toContain("<meta data-hf-id");
+  });
+
+  it("is idempotent: a second call mints nothing and is byte-stable", () => {
+    const html = `<!doctype html><html><body><div><p>a</p></div></body></html>`;
+    const once = ensureHfIds(html);
+    const twice = ensureHfIds(once);
+    expect(twice).toBe(once);
+  });
+
+  it("pins existing data-hf-id and mints around it", () => {
+    const html = `<!doctype html><html><body>
+      <div data-hf-id="hf-keep"><p>a</p></div></body></html>`;
+    const out = ensureHfIds(html);
+    expect(out).toContain('data-hf-id="hf-keep"');
+    expect(ids(out)).toContain("hf-keep");
+    expect(ids(out)).toHaveLength(2); // div pinned + p minted
+  });
+
+  it("two identical sibling nodes get distinct ids", () => {
+    const html = `<!doctype html><html><body>
+      <p class="x">same</p><p class="x">same</p></body></html>`;
+    const got = ids(ensureHfIds(html));
+    expect(new Set(got).size).toBe(got.length);
+  });
+
+  it("is deterministic: same input → same ids", () => {
+    const html = `<!doctype html><html><body><div><p>a</p><span>b</span></div></body></html>`;
+    expect(ids(ensureHfIds(html))).toEqual(ids(ensureHfIds(html)));
+  });
+
+  it("mintHfId rehashes on collision against the assigned set", () => {
+    const { document } = parseHTML(`<p class="x">same</p>`);
+    const el = document.querySelector("p") as Element;
+    const assigned = new Set<string>();
+    const a = mintHfId(el, assigned);
+    const b = mintHfId(el, assigned); // identical element, same assigned set
+    expect(a).not.toBe(b);
+    expect(a).toMatch(/^hf-[a-z0-9]{4}$/);
+    expect(b).toMatch(/^hf-[a-z0-9]{4}$/);
+  });
+});
+
+// Lock the edit-lifecycle behavior. These pin BOTH the guarantee that holds
+// once ids are persisted to source (pinning) AND the two limitations that hold
+// while they are not (design §3 write-back is not yet wired — see
+// notes/r1-stable-hf-ids-design.md "Implementation status & verified lifecycle gap").
+describe("ensureHfIds — edit lifecycle (R1 stability)", () => {
+  it("pinned id survives a content edit (the §3 write-back guarantee)", () => {
+    // Element already carries data-hf-id in source (as it would after write-back).
+    const edited = doc(`<p class="body" data-hf-id="hf-abcd">Hello world</p>`);
+    expect(idOf(ensureHfIds(edited), "p.body")).toBe("hf-abcd");
+  });
+
+  it("KNOWN LIMITATION: an unpinned id changes when the element's text is edited", () => {
+    // No data-hf-id in source → every parse re-mints from content. Editing the
+    // text changes the hash, so the id drifts. This is the "pure-hash" mode the
+    // design rejected; flip this assertion to .toBe once write-back lands.
+    const before = idOf(ensureHfIds(doc(`<p class="body">Hello</p>`)), "p.body");
+    const after = idOf(ensureHfIds(doc(`<p class="body">Hello world</p>`)), "p.body");
+    expect(before).not.toBe(after);
+  });
+
+  it("KNOWN LIMITATION: an unpinned id changes when an attribute is edited", () => {
+    const before = idOf(ensureHfIds(doc(`<p class="body">x</p>`)), "p");
+    const after = idOf(ensureHfIds(doc(`<p class="lead">x</p>`)), "p");
+    expect(before).not.toBe(after);
+  });
+
+  it("KNOWN LIMITATION: identical-content siblings have no content-stable id for the 2nd occurrence", () => {
+    // Insertion stability holds for DISTINCT content (covered elsewhere), but a
+    // second identical sibling collides and gets a position-derived dedup id —
+    // there is no content-stable handle for it. The first keeps the base id.
+    const single = idOf(ensureHfIds(doc(`<p class="x">same</p>`)), "p.x");
+    const pair = ids(ensureHfIds(doc(`<p class="x">same</p><p class="x">same</p>`)));
+    expect(pair[0]).toBe(single); // first identical element: stable, content-derived
+    expect(pair[1]).not.toBe(single); // second: dedup id, exists only by position
+  });
+});
diff --git a/packages/core/src/parsers/hfIds.ts b/packages/core/src/parsers/hfIds.ts
@@ -0,0 +1,104 @@
+/**
+ * Stable hf- element id minting (R1). Node-safe (linkedom only, not browser DOM).
+ *
+ * Two surfaces share these helpers:
+ *  - ensureHfIds(html): node-id surface — mints data-hf-id on every element.
+ *  - mintHfId(el, assigned): shared by htmlParser for clip ids.
+ *
+ * Hash is CONTENT ONLY (tag + sorted attrs + own text) — no sibling position,
+ * so inserting a non-identical sibling never shifts another element's id.
+ */
+import { parseHTML } from "linkedom";
+
+// Non-editable / non-visual elements that should never receive a stable id.
+const EXCLUDED_TAGS = new Set(["script", "style", "template", "meta", "link", "noscript", "base"]);
+
+// 32-bit FNV-1a. Pure, deterministic, no crypto, no Math.random.
+function fnv1a(str: string): number {
+  let h = 0x811c9dc5;
+  for (let i = 0; i < str.length; i++) {
+    h ^= str.charCodeAt(i);
+    h = Math.imul(h, 0x01000193);
+  }
+  return h >>> 0;
+}
+
+// 4 base-36 chars · 36^4 ≈ 1.68M ids per document. Birthday-paradox collision
+// ≈ N²/(2·36^4): well under 1% per document after dup rehash at realistic
+// clip-model sizes (≤ a few hundred elements). The dup-rehash in mintHfId
+// resolves the rare collision; width is deliberately small for readable ids.
+function toHfId(hash: number): string {
+  const s = (hash >>> 0).toString(36);
+  // Use suffix (most-avalanched bits) for better distribution within the 4-char window.
+  const four = s.length >= 4 ? s.slice(-4) : s.padStart(4, "0");
+  return `hf-${four}`;
+}
+
+// Element's own direct text (TEXT_NODE children), not descendants'.
+function ownText(el: Element): string {
+  let text = "";
+  el.childNodes.forEach((n) => {
+    if (n.nodeType === 3) text += (n as Text).nodeValue ?? "";
+  });
+  return text.trim();
+}
+
+function contentKey(el: Element): string {
+  // Exclude all data-hf-* attrs (ids, studio state) — they must not influence the hash.
+  // Use \x00 / \x01 separators (invalid in HTML attrs) to prevent ambiguous serialization.
+  const attrs = Array.from(el.attributes)
+    .filter((a) => !a.name.startsWith("data-hf-"))
+    .map((a) => `${a.name}\x00${a.value}`)
+    .sort()
+    .join("\x01");
+  return `${el.tagName.toLowerCase()}|${attrs}|${ownText(el)}`;
+}
+
+export function mintHfId(el: Element, assigned: Set<string>): string {
+  const key = contentKey(el);
+  let id = toHfId(fnv1a(key));
+  let dup = 0;
+  while (assigned.has(id)) {
+    dup += 1;
+    // Graceful fallback instead of a hard throw: rehashing only fails to find a
+    // free 4-char slot in a pathological document (~1.6M identical elements).
+    // Rather than crash the whole parse, widen the id with the dup counter —
+    // still deterministic and unique, just longer than the 4-char norm.
+    if (dup > 10000) {
+      id = `hf-${(fnv1a(key) >>> 0).toString(36)}-${dup}`;
+      break;
+    }
+    id = toHfId(fnv1a(`${key}#${dup}`));
+  }
+  assigned.add(id);
+  return id;
+}
+
+export function ensureHfIds(html: string): string {
+  // Mirror parseSourceDocument's fragment-wrapping so bare fragments don't land
+  // outside <body> in linkedom, which would cause body.querySelectorAll to return [].
+  const hasDocumentShell = /<!doctype|<html[\s>]/i.test(html);
+  const wrapped = !hasDocumentShell;
+  const { document } = wrapped
+    ? parseHTML(`<!DOCTYPE html><html><head></head><body>${html}</body></html>`)
+    : parseHTML(html);
+  const body = document.body;
+  if (!body) return html;
+
+  const assigned = new Set<string>();
+  // Seed with already-present ids (pin) so fresh mints never collide with them.
+  // Scope to <body> to match the mint walk below — a stray data-hf-id in <head>
+  // must not pin an id into the set that a body element would then be bumped off.
+  for (const el of Array.from(body.querySelectorAll("[data-hf-id]"))) {
+    const existing = el.getAttribute("data-hf-id");
+    if (existing) assigned.add(existing);
+  }
+
+  for (const el of Array.from(body.querySelectorAll("*"))) {
+    if (EXCLUDED_TAGS.has(el.tagName.toLowerCase())) continue;
+    if (el.getAttribute("data-hf-id")) continue; // pinned
+    el.setAttribute("data-hf-id", mintHfId(el, assigned));
+  }
+
+  return wrapped ? document.body.innerHTML || "" : document.toString();
+}