feat: headless Chrome fallback for browser web search outside broadcast

browse-search.mjs only connected to the on-screen broadcast Chrome over
CDP, so browser-based Google search worked only during a live broadcast;
plain text turns fell through to the DDG cascade. Add a headless fallback
(system Chrome via channel:'chrome', else Playwright's bundled chromium)
for `search` mode so general conversation can use Google at no API cost.
`youtube` still requires the visible broadcast Chrome.

Detect Google's /sorry bot-detection interstitial structurally by URL and
fail fast so the caller fail-opens to DDG/Brave/Wikipedia instead of
treating the challenge page as empty results.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

bot/scripts/stream-test/browse-search.mjs

@@ -1,11 +1,17 @@
-// True-mode browser action core. Drives the on-screen Chrome (CDP at CDP_PORT,
-// default 9222) so the action is visible on the Go-Live broadcast, and prints a
-// JSON result on stdout for the Python `browseAndSearch` tool to wrap.
+// Browser action core. Prefers the on-screen Chrome (CDP at CDP_PORT, default
+// 9222) so the action is visible on the Go-Live broadcast, and prints a JSON
+// result on stdout for the Python `browseAndSearch` tool to wrap.
 //
 //   node browse-search.mjs "<query>" [search|youtube]
 //
 // - search  : Google-search the query, return the top organic results.
 // - youtube : search YouTube and play the first result.
+//
+// When no broadcast Chrome is reachable on CDP (e.g. a plain text conversation
+// with no active broadcast), `search` falls back to a headless Chromium we
+// launch ourselves, so browser-based Google search still works at zero API
+// cost. `youtube` only makes sense on the visible broadcast Chrome, so it never
+// uses the headless fallback.
 import { chromium } from 'playwright';
 
 const CDP = process.env.CDP_PORT || '9222';
@@ -18,11 +24,48 @@ const out = (o) => { process.stdout.write(JSON.stringify(o)); };
 
 if (!query) { out({ ok: false, error: 'no query' }); process.exit(1); }
 
-let b;
+let browser;          // playwright Browser: connected (CDP) or launched (headless)
+let launched = false; // true when we launched our own headless browser
+let page;
+
+// Acquire a page from the broadcast Chrome over CDP, or — for search mode only —
+// from a headless Chromium we launch when no broadcast Chrome is reachable.
+async function acquirePage() {
+  try {
+    browser = await chromium.connectOverCDP(`http://${CDP_HOST}:${CDP}`);
+    const ctx = browser.contexts()[0];
+    page = ctx.pages()[0] || (await ctx.newPage());
+    return;
+  } catch (e) {
+    // YouTube playback must be on the visible broadcast Chrome; no fallback.
+    if (mode === 'youtube') throw e;
+  }
+  // Prefer the system Chrome the host already has (channel: 'chrome') so the
+  // fallback needs no extra Playwright browser download; degrade to Playwright's
+  // bundled chromium if no system Chrome is installed.
+  launched = true;
+  let launchErr;
+  for (const opts of [{ headless: true, channel: 'chrome' }, { headless: true }]) {
+    try {
+      browser = await chromium.launch(opts);
+      break;
+    } catch (e) {
+      launchErr = e;
+      browser = undefined;
+    }
+  }
+  if (!browser) throw launchErr;
+  const ctx = await browser.newContext({
+    locale: 'ko-KR',
+    userAgent:
+      'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 ' +
+      '(KHTML, like Gecko) Chrome/148.0.0.0 Safari/537.36',
+  });
+  page = await ctx.newPage();
+}
+
 try {
-  b = await chromium.connectOverCDP(`http://${CDP_HOST}:${CDP}`);
-  const ctx = b.contexts()[0];
-  const page = ctx.pages()[0] || (await ctx.newPage());
+  await acquirePage();
   page.setDefaultTimeout(20000);
   await page.bringToFront().catch(() => {});
 
@@ -38,6 +81,15 @@ try {
   } else {
     await page.goto(`https://www.google.com/search?q=${encodeURIComponent(query)}&hl=ko`, { waitUntil: 'domcontentloaded' });
     await page.waitForTimeout(1500);
+    // Google serves its bot-detection interstitial (/sorry/index) to sessions it
+    // suspects are automated. Detect it structurally (by URL, locale-independent)
+    // and fail fast so the Python caller fail-opens to the DDG cascade instead of
+    // treating an empty challenge page as "no results".
+    if (page.url().includes('/sorry/')) {
+      await browser.close();
+      out({ ok: false, error: 'google-bot-challenge', headless: launched });
+      process.exit(1);
+    }
     const results = await page.evaluate(() => {
       const seen = new Set();
       const items = [];
@@ -55,11 +107,11 @@ try {
       }
       return items;
     });
-    out({ ok: true, mode, query, count: results.length, results });
+    out({ ok: true, mode, query, count: results.length, results, headless: launched });
   }
-  await b.close();
+  await browser.close();
 } catch (e) {
-  try { await b?.close(); } catch { /* ignore */ }
+  try { await browser?.close(); } catch { /* ignore */ }
   out({ ok: false, error: String(e?.message || e) });
   process.exit(1);
 }

src/jarvis/tools/builtin/web_search.spec.md

@@ -22,7 +22,17 @@ path (evals, text entry) and falls back to the master flag:
 
 - **on-screen Chrome**: `browser_search()` drives Chrome (Node CDP helper
   `bot/scripts/stream-test/browse-search.mjs`) to Google-search the query, so
-  the action is visible on the Go-Live broadcast.
+  the action is visible on the Go-Live broadcast. When no broadcast Chrome is
+  reachable on CDP (e.g. a plain text turn with no active broadcast), the helper
+  falls back to launching its own headless Chrome (system `channel: 'chrome'`,
+  else Playwright's bundled chromium) so browser-based Google search still works
+  with no API cost. The `youtube` action never uses the headless fallback (it
+  only makes sense on the visible broadcast Chrome). Caveat: a freshly-launched
+  headless session with no logged-in profile can be served Google's
+  bot-detection interstitial (`/sorry/index`); the helper detects this
+  structurally by URL and fails fast, so the caller fail-opens to the DDG /
+  Brave / Wikipedia cascade rather than treating the challenge page as "no
+  results".
 - **Gemini**: answers, with the sub-mode chosen by `cfg.gemini_auth`
   (env `GEMINI_AUTH`, default `oauth`):
   - `oauth` (default): `gemini_cli_search()` shells out to the Gemini CLI