fix(server): PR watch wakes quote a bot comment's text, not its HTML

apps/server/src/orchestration-v2/pullRequestWatch.test.ts

@@ -735,6 +735,47 @@ describe("evaluatePullRequestWatch", () => {
     assert.include(text, `"${"x".repeat(196)}..."`);
   });
 
+  it.each([
+    [
+      "a bot's leading badge",
+      '<a href="https://app.coderabbit.ai/change-stack/o/r/pull/1?cs_source=review_comment"><picture><source media="(prefers-color-scheme: dark)" srcset="https://coderabbit.ai/badge-dark.svg"><img alt="Severity > `Minor`" src="https://coderabbit.ai/badge.svg"></picture></a>\r\n\r\n_⚠️ Potential issue_ | _🟠 Major_\r\n\r\n**Guard the empty list.**',
+      "_⚠️ Potential issue_ | _🟠 Major_ **Guard the empty list.**",
+    ],
+    [
+      "tags in prose and a hidden comment",
+      "<!-- state -->\n<details>\n<summary>Why</summary>\n\n<p>Guard this</p>, then <br/>retry.\n</details>",
+      "Why Guard this , then retry.",
+    ],
+    [
+      "type arguments",
+      "Return Promise<void>, not Promise<A> or Vec<code>.",
+      "Return Promise<void>, not Promise<A> or Vec<code>.",
+    ],
+    [
+      "code spans and fences",
+      "Replace `<div>` with ``<span>``:\n~~~html\n<p>Done</p>\n~~~\r\nor\r\n```html\r\n<b>Done</b>\r\n```",
+      "Replace `<div>` with ``<span>``: ~~~html <p>Done</p> ~~~ or ```html <b>Done</b> ```",
+    ],
+    [
+      "indented code",
+      "Use this replacement:\n\n    <div>\n      <span>Ready</span>\n    </div>\n\nKeep the <b>wrapper</b>.",
+      "Use this replacement: <div> <span>Ready</span> </div> Keep the wrapper .",
+    ],
+    ["an indented continuation of an HTML block", "<div>\n    <span>Ready</span>\n</div>", "Ready"],
+  ])("quotes a remark's text without its HTML tags: %s", (_name, body, quoted) => {
+    const report = evaluatePullRequestWatch(watch(), detail(), [
+      remark("reviewer", "2026-10-02T12:06:00Z", body),
+    ]);
+    const { text } = pullRequestWatchMessage({
+      number: 1,
+      url: "https://github.com/o/r/pull/1",
+      baseBranch: "main",
+      headSha: null,
+      report,
+    });
+    assert.include(text, `: "${quoted}" `);
+  });
+
   it("reports a conflict once, until the branch is clean again", () => {
     const conflicting = detail({ mergeability: "conflicting" });
     const first = evaluatePullRequestWatch(watch(), conflicting, noRemarks);

apps/server/src/orchestration-v2/pullRequestWatch.ts

@@ -261,9 +261,81 @@ export function evaluatePullRequestWatch(
   };
 }
 
+// GitHub renders these, so in a remark's prose they are presentation, not what the reviewer said.
+// Lowercase only: `Promise<T>` is a type argument.
+const TAG_NAMES =
+  "a|abbr|b|blockquote|br|caption|cite|code|dd|del|details|div|dl|dt|em|h[1-6]|hr|i|img|ins|kbd|li|mark|ol|p|picture|pre|q|s|samp|small|source|span|strike|strong|sub|summary|sup|table|tbody|td|tfoot|th|thead|tr|tt|ul|var";
+// A code span comes first so its tags stay literal; it cannot cross a blank line. A tag stops at
+// the next `<`, which keeps unclosed openers from rescanning the rest of the body.
+const PROSE_MARKUP = new RegExp(
+  "(?<!`)(?<ticks>`+)(?!`)(?:(?!\\n[ \\t\\r]*\\n)[\\s\\S])*?(?<!`)\\k<ticks>(?!`)" +
+    `|<(?<closing>/?)(?<name>${TAG_NAMES})(?=[\\s/>])(?:[^<>"']|"[^"]*"|'[^']*')*>`,
+  "g",
+);
+
+function withoutTags(prose: string): string {
+  // An opener glued to an identifier is a type argument (`Vec<code>`); it and its closer stay.
+  const kept = new Map<string, number>();
+  let text = "";
+  let from = 0;
+  for (const match of prose.matchAll(PROSE_MARKUP)) {
+    const name = match.groups?.name;
+    if (name === undefined) continue;
+    const held = kept.get(name) ?? 0;
+    if (match.groups?.closing === "/") {
+      if (held > 0) {
+        kept.set(name, held - 1);
+        continue;
+      }
+    } else if (/\w/.test(prose.charAt(match.index - 1))) {
+      kept.set(name, held + 1);
+      continue;
+    }
+    text += `${prose.slice(from, match.index)} `;
+    from = match.index + match[0].length;
+  }
+  return text + prose.slice(from);
+}
+
+// Drops HTML tags from Markdown prose. Fenced and indented code blocks pass through untouched.
+function withoutMarkup(body: string): string {
+  const parts: Array<string> = [];
+  let prose: Array<string> = [];
+  let fence: string | undefined;
+  let afterText = false;
+  const code = (line: string) => {
+    if (prose.length > 0) parts.push(withoutTags(prose.join("\n")));
+    prose = [];
+    parts.push(line);
+    afterText = false;
+  };
+  for (const line of body.split("\n")) {
+    if (fence !== undefined) {
+      const closer = line.trim();
+      if (closer.length >= fence.length && closer.replaceAll(fence.charAt(0), "") === "") {
+        fence = undefined;
+      }
+      code(line);
+      continue;
+    }
+    const opener = /^ {0,3}(`{3,}|~{3,})([\s\S]*)$/.exec(line);
+    if (opener?.[1] !== undefined && !(opener[1].startsWith("`") && opener[2]?.includes("`"))) {
+      fence = opener[1];
+      code(line);
+    } else if (!afterText && /^(?: {4}| {0,3}\t)\s*\S/.test(line)) {
+      // Indented code needs a blank line before it; after text the indent is a continuation.
+      code(line);
+    } else {
+      prose.push(line);
+      afterText = line.trim() !== "";
+    }
+  }
+  if (prose.length > 0) parts.push(withoutTags(prose.join("\n")));
+  return parts.join("\n");
+}
+
 function snippet(body: string): string {
-  const text = body
-    .replaceAll(/<!--[\s\S]*?-->/g, " ")
+  const text = withoutMarkup(body.replaceAll(/<!--[\s\S]*?-->/g, " "))
     .replaceAll(/\s+/g, " ")
     .trim();
   if (text.length <= SNIPPET_LENGTH) return text;