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

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

@@ -735,6 +735,58 @@ describe("evaluatePullRequestWatch", () => {
     assert.include(text, `"${"x".repeat(196)}..."`);
   });
 
+  it("quotes a remark's text instead of its leading HTML, and keeps tags in prose and code", () => {
+    const 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>';
+    const report = evaluatePullRequestWatch(watch(), detail(), [
+      remark(
+        "coderabbitai",
+        "2026-10-02T12:06:00Z",
+        `${badge}\n\n_⚠️ Potential issue_ | _🟠 Major_\n\n**Guard the empty list.**`,
+      ),
+      remark(
+        "reviewer",
+        "2026-10-02T12:07:00Z",
+        "Return Promise<void>, not Promise<A> or Promise<a>.",
+      ),
+      remark(
+        "designer",
+        "2026-10-02T12:08:00Z",
+        "Replace `<div>` with ``<span>``:\n~~~html\n<p>Done</p>\n~~~",
+      ),
+      remark("maintainer", "2026-10-02T12:09:00Z", "<p>Guard this</p>, and `<!--` stays escaped."),
+      remark(
+        "summary-bot",
+        "2026-10-02T12:10:00Z",
+        `<!-- ${"x".repeat(3_000)} -->\nCheck the retry.`,
+      ),
+      remark(
+        "fencer",
+        "2026-10-02T12:12:00Z",
+        "Use this literal opener:\n```html\n<!--\n```\nGuard the empty list before indexing it.",
+      ),
+      remark("spacer", "2026-10-02T12:11:00Z", `${"\n".repeat(2_001)}Guard the empty list.`),
+    ]);
+    const { text } = pullRequestWatchMessage({
+      number: 1,
+      url: "https://github.com/o/r/pull/1",
+      baseBranch: "main",
+      headSha: null,
+      report,
+    });
+    assert.include(text, `"_⚠️ Potential issue_ | _🟠 Major_ **Guard the empty list.**"`);
+    assert.include(text, `"Return Promise<void>, not Promise<A> or Promise<a>."`);
+    assert.include(text, '"Replace `<div>` with ``<span>``: ~~~html <p>Done</p> ~~~"');
+    assert.include(text, 'maintainer on src/index.ts: "Guard this , and `<!--` stays escaped."');
+    assert.include(text, '"Check the retry."');
+    assert.include(
+      text,
+      '"Use this literal opener: ```html <!-- ``` Guard the empty list before indexing it."',
+    );
+    assert.include(text, 'spacer on src/index.ts: "Guard the empty list."');
+    assert.notInclude(text, "<picture");
+  });
+
   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

@@ -13,6 +13,15 @@ export const PULL_REQUEST_WATCH_WAKE_LIMIT = 10;
 export const PULL_REQUEST_WATCH_NO_CHECKS_MS = 15 * 60_000;
 const LISTED_ITEMS = 10;
 const SNIPPET_LENGTH = 200;
+// Remark bodies are untrusted and run to 64 KB, and unclosed openers make the markup pass
+// backtrack, so it only reads far enough to fill an excerpt once HTML comments are gone.
+const SNIPPET_SCAN_LENGTH = 2_000;
+// One left-to-right pass keeps code spans' tags and reads attribute backticks as part of the tag.
+// A lookahead pins each delimiter run at full length, so a long run can't backtrack cubically.
+// Element names are lowercase. A one-letter element needs attributes or a later closing tag,
+// so Promise<a> stays.
+const MARKUP =
+  /(?<code>(?<!`)(?=(?<ticks>`+))\k<ticks>[\s\S]*?(?<!`)\k<ticks>(?!`)|(?<!~)(?=(?<tildes>~{3,}))\k<tildes>[\s\S]*?\k<tildes>)|<(?<one>[abip])>(?=[\s\S]*?<\/\k<one>>)|<\/(?:a|b|blockquote|br|code|dd|details|div|dl|dt|em|h[1-6]|hr|i|img|li|ol|p|picture|pre|source|span|strong|sub|summary|sup|table|tbody|td|th|thead|tr|ul)\s*>|<(?:blockquote|br|code|dd|details|div|dl|dt|em|h[1-6]|hr|img|li|ol|picture|pre|source|span|strong|sub|summary|sup|table|tbody|td|th|thead|tr|ul)(?=[\s/>])(?:[^>"']|"[^"]*"|'[^']*')*>|<(?:a|b|i|p)\s(?:[^>"']|"[^"]*"|'[^']*')*>/g;
 
 export type PullRequestWatchChange =
   | { readonly kind: "checks-failed"; readonly failed: ReadonlyArray<PullRequestCheck> }
@@ -261,9 +270,25 @@ export function evaluatePullRequestWatch(
   };
 }
 
+// indexOf keeps this linear on unclosed openers, so it can read the whole body.
+function withoutHtmlComments(body: string): string {
+  let text = "";
+  let from = 0;
+  for (let open = body.indexOf("<!--"); open !== -1; open = body.indexOf("<!--", from)) {
+    const close = body.indexOf("-->", open + 4);
+    if (close === -1) break;
+    text += `${body.slice(from, open)} `;
+    from = close + 3;
+  }
+  return text + body.slice(from);
+}
+
 function snippet(body: string): string {
-  const text = body
-    .replaceAll(/<!--[\s\S]*?-->/g, " ")
+  const text = withoutHtmlComments(body)
+    .replaceAll(/\s+/g, " ")
+    .slice(0, SNIPPET_SCAN_LENGTH)
+    .toWellFormed()
+    .replaceAll(MARKUP, (match, code: string | undefined) => (code === undefined ? " " : match))
     .replaceAll(/\s+/g, " ")
     .trim();
   if (text.length <= SNIPPET_LENGTH) return text;