@apmantza/greedysearch-pi 1.4.1 → 1.4.2

package/README.md

@@ -1,208 +1,219 @@
-# GreedySearch for Pi
-
-Pi extension that adds a `greedy_search` tool — fans out queries to Perplexity, Bing Copilot, and Google AI simultaneously and returns AI-synthesized answers with deduped sources. Streams progress as each engine completes.
-
-Forked from [GreedySearch-claude](https://github.com/apmantza/GreedySearch-claude).
-
-## What's New (v1.4.1)
-
-- **Fixed parallel synthesis** — multiple `greedy_search` calls with `synthesize: true` now run safely in parallel. Each search creates a fresh Gemini tab that gets cleaned up after synthesis, preventing tab conflicts and "Uncaught" errors.
-
-## What's New (v1.4.0)
-
-- **Grounded synthesis** — Gemini now receives a normalized source registry with stable source IDs, agreement summaries, caveats, and cited claims
-- **Real deep research** — top sources are fetched before synthesis so deep research answers are grounded in fetched evidence, not just engine summaries
-- **Richer source metadata** — source output now includes canonical URLs, domains, source types, per-engine attribution, and confidence metadata
-- **Cleaner tab lifecycle** — temporary Perplexity, Bing, and Google tabs are closed after each fan-out search, and synthesis finishes on the Gemini tab
-- **Isolated Chrome targeting** — GreedySearch now refuses to fall back to your normal Chrome session, preventing stray remote-debugging prompts
-
-## Install
-
-```bash
-pi install npm:@apmantza/greedysearch-pi
-```
-
-Or directly from git:
-
-```bash
-pi install git:github.com/apmantza/GreedySearch-pi
-```
-
-## Quick Start
-
-Once installed, Pi gains a `greedy_search` tool. The model will use it automatically for questions about current libraries, error messages, version-specific docs, etc.
-
-```
-greedy_search({ query: "What's new in React 19?", engine: "all" })
-```
-
-## Parameters
-
-| Parameter | Type | Default | Description |
-|-----------|------|---------|-------------|
-| `query` | string | required | The search question |
-| `engine` | string | `"all"` | Engine to use (see below) |
-| `synthesize` | boolean | `false` | Synthesize results into one answer via Gemini |
-| `fullAnswer` | boolean | `false` | Return complete answer (~3000+ chars) vs truncated preview (~300 chars) |
-
-## Engines
-
-| Engine | Alias | Latency | Best for |
-|--------|-------|---------|----------|
-| `all` | — | 30-90s | Highest confidence — all 3 engines in parallel (default) |
-| `perplexity` | `p` | 15-30s | Technical Q&A, code explanations, documentation |
-| `bing` | `b` | 15-30s | Recent news, Microsoft ecosystem |
-| `google` | `g` | 15-30s | Broad coverage, multiple perspectives |
-| `gemini` | `gem` | 15-30s | Google's AI with different training data |
-
-## Streaming Progress
-
-When using `engine: "all"`, the tool streams progress as each engine completes:
-
-```
-**Searching...** ⏳ perplexity · ⏳ bing · ⏳ google
-**Searching...** ✅ perplexity done · ⏳ bing · ⏳ google
-**Searching...** ✅ perplexity done · ✅ bing done · ⏳ google
-**Searching...** ✅ perplexity done · ✅ bing done · ✅ google done
-```
-
-## Synthesis Mode
-
-For complex research questions, use `synthesize: true` with `engine: "all"`:
-
-```
-greedy_search({ query: "best auth patterns for SaaS in 2026", engine: "all", synthesize: true })
-```
-
-This deduplicates sources across engines, builds a normalized source registry, and feeds that context to Gemini for one clean synthesized answer. Adds ~30s but now returns agreement summaries, caveats, key claims, and better-labeled top sources.
-
-For the most grounded mode, use deep research from the CLI:
-
-```bash
-node search.mjs all "best auth patterns for SaaS in 2026" --deep-research
-```
-
-Deep research fetches top source pages before synthesis and reports source confidence metadata such as agreement level, fetched-source success rate, and source mix.
-
-**Use synthesis when:**
-- You need one definitive answer, not multiple perspectives
-- You're researching a topic to write about or make a decision
-- Token efficiency matters (one answer vs three)
-
-**Skip synthesis when:**
-- You want to see where engines disagree
-- Speed matters
-
-## Full vs Short Answers
-
-Default mode returns ~300 char summaries to save tokens. Use `fullAnswer: true` for complete responses:
-
-```
-greedy_search({ query: "explain the React compiler", engine: "perplexity", fullAnswer: true })
-```
-
-## Examples
-
-**Quick technical lookup:**
-```
-greedy_search({ query: "How to use async await in Python", engine: "perplexity" })
-```
-
-**Compare tools (see where engines agree/disagree):**
-```
-greedy_search({ query: "Prisma vs Drizzle in 2026", engine: "all" })
-```
-
-**Research with synthesis:**
-```
-greedy_search({ query: "Best practices for monorepo structure", engine: "all", synthesize: true })
-```
-
-**Debug an error:**
-```
-greedy_search({ query: "Error: Cannot find module 'react-dom/client' Next.js 15", engine: "all" })
-```
-
-## Requirements
-
-- **Chrome** — must be installed. The extension auto-launches a dedicated Chrome instance on port 9222 with its own isolated profile and DevTools port file, separate from your main browser session.
-- **Node.js 22+** — for built-in `fetch` and WebSocket support.
-
-## Setup (first time)
-
-To pre-launch the dedicated GreedySearch Chrome instance:
-
-```bash
-node ~/.pi/agent/git/GreedySearch-pi/launch.mjs
-```
-
-Stop it when done:
-
-```bash
-node ~/.pi/agent/git/GreedySearch-pi/launch.mjs --kill
-```
-
-Check status:
-
-```bash
-node ~/.pi/agent/git/GreedySearch-pi/launch.mjs --status
-```
-
-## Testing
-
-Run the test suite to verify everything works:
-
-```bash
-./test.sh           # full suite (~3-4 min)
-./test.sh quick     # skip parallel tests (~1 min)
-./test.sh parallel  # parallel race condition tests only
-```
-
-Tests verify:
-- Single engine mode (perplexity, bing, google)
-- Sequential "all" mode searches
-- Parallel "all" mode (5 concurrent searches) — detects tab race conditions
-- Synthesis mode with Gemini
-
-## Troubleshooting
-
-### "Chrome not found"
-Set the path explicitly:
-```bash
-export CHROME_PATH="/path/to/chrome"
-```
-
-### "CDP timeout" or "Chrome may have crashed"
-Restart GreedySearch Chrome:
-```bash
-node ~/.pi/agent/git/GreedySearch-pi/launch.mjs --kill
-node ~/.pi/agent/git/GreedySearch-pi/launch.mjs
-```
-
-### Google / Bing "verify you're human"
-The extension auto-clicks verification buttons and Cloudflare Turnstile challenges. For hard CAPTCHAs (image puzzles), solve manually in the Chrome window that opens.
-
-### Parallel searches failing
-Earlier versions shared Chrome tabs between concurrent searches, causing `ERR_ABORTED` errors. Version 1.2.0+ creates fresh tabs for each search, allowing safe parallel execution.
-
-### Search hangs
-Chrome may be unresponsive. Restart it with `launch.mjs --kill` then `launch.mjs`.
-
-### Sources are junk links
-This was a known issue with Gemini sources. If you're on an older version, update:
-```bash
-pi install npm:@apmantza/greedysearch-pi
-```
-
-## How It Works
-
-- `index.ts` — Pi extension, registers `greedy_search` tool with streaming progress
-- `search.mjs` — CLI runner, spawns extractors in parallel, emits `PROGRESS:` events to stderr
-- `launch.mjs` — launches dedicated Chrome on port 9222 with isolated profile
-- `extractors/` — per-engine CDP scrapers (Perplexity, Bing Copilot, Google AI, Gemini)
-- `cdp.mjs` — Chrome DevTools Protocol CLI for browser automation
-- `skills/greedy-search/SKILL.md` — skill file that guides the model on when/how to use greedy_search
-
-## License
-
-MIT
+# GreedySearch for Pi
+
+Pi extension that adds a `greedy_search` tool — fans out queries to Perplexity, Bing Copilot, and Google AI simultaneously and returns AI-synthesized answers with deduped sources. Streams progress as each engine completes.
+
+Forked from [GreedySearch-claude](https://github.com/apmantza/GreedySearch-claude).
+
+## Install
+
+```bash
+pi install npm:@apmantza/greedysearch-pi
+```
+
+Or directly from git:
+
+```bash
+pi install git:github.com/apmantza/GreedySearch-pi
+```
+
+## Quick Start
+
+Once installed, Pi gains a `greedy_search` tool. The model will use it automatically for questions about current libraries, error messages, version-specific docs, etc.
+
+```
+greedy_search({ query: "What's new in React 19?", engine: "all" })
+```
+
+## Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `query` | string | required | The search question |
+| `engine` | string | `"all"` | Engine to use (see below) |
+| `synthesize` | boolean | `false` | Synthesize results into one answer via Gemini |
+| `fullAnswer` | boolean | `false` | Return complete answer (~3000+ chars) vs truncated preview (~300 chars) |
+
+## Engines
+
+| Engine | Alias | Latency | Best for |
+|--------|-------|---------|----------|
+| `all` | — | 30-90s | Highest confidence — all 3 engines in parallel (default) |
+| `perplexity` | `p` | 15-30s | Technical Q&A, code explanations, documentation |
+| `bing` | `b` | 15-30s | Recent news, Microsoft ecosystem |
+| `google` | `g` | 15-30s | Broad coverage, multiple perspectives |
+| `gemini` | `gem` | 15-30s | Google's AI with different training data |
+
+## Streaming Progress
+
+When using `engine: "all"`, the tool streams progress as each engine completes:
+
+```
+**Searching...** ⏳ perplexity · ⏳ bing · ⏳ google
+**Searching...** ✅ perplexity done · ⏳ bing · ⏳ google
+**Searching...** ✅ perplexity done · ✅ bing done · ⏳ google
+**Searching...** ✅ perplexity done · ✅ bing done · ✅ google done
+```
+
+## Synthesis Mode
+
+For complex research questions, use `synthesize: true` with `engine: "all"`:
+
+```
+greedy_search({ query: "best auth patterns for SaaS in 2026", engine: "all", synthesize: true })
+```
+
+This deduplicates sources across engines, builds a normalized source registry, and feeds that context to Gemini for one clean synthesized answer. Adds ~30s but returns agreement summaries, caveats, key claims, and better-labeled top sources.
+
+**Use synthesis when:**
+- You need one definitive answer, not multiple perspectives
+- You're researching a topic to write about or make a decision
+- Token efficiency matters (one answer vs three)
+
+**Skip synthesis when:**
+- You want to see where engines disagree
+- Speed matters
+
+## Full vs Short Answers
+
+Default mode returns ~300 char summaries to save tokens. Use `fullAnswer: true` for complete responses:
+
+```
+greedy_search({ query: "explain the React compiler", engine: "perplexity", fullAnswer: true })
+```
+
+## Examples
+
+**Quick technical lookup:**
+
+```
+greedy_search({ query: "How to use async await in Python", engine: "perplexity" })
+```
+
+**Compare tools (see where engines agree/disagree):**
+
+```
+greedy_search({ query: "Prisma vs Drizzle in 2026", engine: "all" })
+```
+
+**Research with synthesis:**
+
+```
+greedy_search({ query: "Best practices for monorepo structure", engine: "all", synthesize: true })
+```
+
+**Debug an error:**
+
+```
+greedy_search({ query: "Error: Cannot find module 'react-dom/client' Next.js 15", engine: "all" })
+```
+
+## Requirements
+
+- **Chrome** — must be installed. The extension auto-launches a dedicated Chrome instance on port 9222 with its own isolated profile and DevTools port file, separate from your main browser session.
+- **Node.js 22+** — for built-in `fetch` and WebSocket support.
+
+## Setup (first time)
+
+To pre-launch the dedicated GreedySearch Chrome instance:
+
+```bash
+node ~/.pi/agent/git/GreedySearch-pi/launch.mjs
+```
+
+Stop it when done:
+
+```bash
+node ~/.pi/agent/git/GreedySearch-pi/launch.mjs --kill
+```
+
+Check status:
+
+```bash
+node ~/.pi/agent/git/GreedySearch-pi/launch.mjs --status
+```
+
+## Testing
+
+Run the test suite to verify everything works:
+
+```bash
+./test.sh           # full suite (~3-4 min)
+./test.sh quick     # skip parallel tests (~1 min)
+./test.sh parallel  # parallel race condition tests only
+```
+
+Tests verify:
+- Single engine mode (perplexity, bing, google)
+- Sequential "all" mode searches
+- Parallel "all" mode (5 concurrent searches) — detects tab race conditions
+- Synthesis mode with Gemini
+
+## Troubleshooting
+
+### "Chrome not found"
+
+Set the path explicitly:
+
+```bash
+export CHROME_PATH="/path/to/chrome"
+```
+
+### "CDP timeout" or "Chrome may have crashed"
+
+Restart GreedySearch Chrome:
+
+```bash
+node ~/.pi/agent/git/GreedySearch-pi/launch.mjs --kill
+node ~/.pi/agent/git/GreedySearch-pi/launch.mjs
+```
+
+### Google / Bing "verify you're human"
+
+The extension auto-clicks verification buttons and Cloudflare Turnstile challenges using broad keyword matching — resilient to variations like "Verify you are human" or localised button text. For hard CAPTCHAs (image puzzles), solve manually in the Chrome window that opens.
+
+### Parallel searches failing
+
+Each search creates a fresh isolated browser tab that is closed after completion, allowing safe parallel execution without tab state conflicts.
+
+### Search hangs
+
+Chrome may be unresponsive. Restart it with `launch.mjs --kill` then `launch.mjs`.
+
+### Sources are empty or junk links
+
+Sources are now extracted by regex-parsing Markdown links (`[title](url)`) from the clipboard text captured after each engine responds — not from DOM selectors that break when the engine's UI updates. If sources are empty, the engine's clipboard copy didn't include formatted links (Bing Copilot currently falls into this category).
+
+## How It Works
+
+- `index.ts` — Pi extension, registers `greedy_search` tool with streaming progress
+- `search.mjs` — CLI runner, spawns extractors in parallel, emits `PROGRESS:` events to stderr
+- `launch.mjs` — launches dedicated Chrome on port 9222 with isolated profile
+- `extractors/` — per-engine CDP scrapers (Perplexity, Bing Copilot, Google AI, Gemini)
+- `cdp.mjs` — Chrome DevTools Protocol CLI for browser automation
+- `skills/greedy-search/SKILL.md` — skill file that guides the model on when/how to use greedy_search
+
+## Changelog
+
+### v1.4.2 (2026-03-25)
+
+- **Fresh isolated tabs** — each search now always creates a new `about:blank` tab via `Target.createTarget` and refreshes the CDP page cache immediately after, preventing SPA navigation failures and stale DOM state from prior queries
+- **Regex-based citation extraction** — all extractors (Perplexity, Bing, Gemini) now parse sources from clipboard Markdown links (`[title](url)`) instead of DOM selectors that break on UI updates
+- **Relaxed verification detection** — `consent.mjs` now uses broad keyword matching (`includes('verify')`, `includes('human')`) instead of anchored regexes, correctly catching button text variants like "Verify you are human" across Cloudflare, Microsoft, and generic modals
+
+---
+
+### v1.4.1
+
+- **Fixed parallel synthesis** — multiple `greedy_search` calls with `synthesize: true` now run safely in parallel. Each search creates a fresh Gemini tab that gets cleaned up after synthesis, preventing tab conflicts and "Uncaught" errors.
+
+### v1.4.0
+
+- **Grounded synthesis** — Gemini now receives a normalized source registry with stable source IDs, agreement summaries, caveats, and cited claims
+- **Real deep research** — top sources are fetched before synthesis so deep research answers are grounded in fetched evidence, not just engine summaries
+- **Richer source metadata** — source output now includes canonical URLs, domains, source types, per-engine attribution, and confidence metadata
+- **Cleaner tab lifecycle** — temporary Perplexity, Bing, and Google tabs are closed after each fan-out search, and synthesis finishes on the Gemini tab
+- **Isolated Chrome targeting** — GreedySearch now refuses to fall back to your normal Chrome session, preventing stray remote-debugging prompts
+
+## License
+
+MIT

package/cdp.mjs

@@ -37,22 +37,22 @@ function getDevToolsActivePortPath() {
   return join(homedir(), '.config', 'google-chrome', 'DevToolsActivePort');
 }
 
-function getWsUrl() {
-  // If CDP_PROFILE_DIR is set (by search.mjs), prefer that profile's port file
-  // so GreedySearch targets its own Chrome, not the user's main session.
-  const profileDir = process.env.CDP_PROFILE_DIR;
-  if (profileDir) {
-    const p = profileDir.replace(/\\/g, '/') + '/DevToolsActivePort';
-    if (existsSync(p)) {
-      const lines = readFileSync(p, 'utf8').trim().split('\n');
-      return `ws://127.0.0.1:${lines[0]}${lines[1]}`;
-    }
-    throw new Error(`GreedySearch DevToolsActivePort not found at ${p}. Refusing to fall back to the main Chrome session.`);
-  }
-  const portFile = getDevToolsActivePortPath();
-  const lines = readFileSync(portFile, 'utf8').trim().split('\n');
-  return `ws://127.0.0.1:${lines[0]}${lines[1]}`;
-}
+function getWsUrl() {
+  // If CDP_PROFILE_DIR is set (by search.mjs), prefer that profile's port file
+  // so GreedySearch targets its own Chrome, not the user's main session.
+  const profileDir = process.env.CDP_PROFILE_DIR;
+  if (profileDir) {
+    const p = profileDir.replace(/\\/g, '/') + '/DevToolsActivePort';
+    if (existsSync(p)) {
+      const lines = readFileSync(p, 'utf8').trim().split('\n');
+      return `ws://127.0.0.1:${lines[0]}${lines[1]}`;
+    }
+    throw new Error(`GreedySearch DevToolsActivePort not found at ${p}. Refusing to fall back to the main Chrome session.`);
+  }
+  const portFile = getDevToolsActivePortPath();
+  const lines = readFileSync(portFile, 'utf8').trim().split('\n');
+  return `ws://127.0.0.1:${lines[0]}${lines[1]}`;
+}
 
 const sleep = (ms) => new Promise(r => setTimeout(r, ms));
 

package/extractors/bing-copilot.mjs

@@ -45,17 +45,14 @@ function cdp(args, timeoutMs = 30000) {
 
 async function getOrOpenTab(tabPrefix) {
   if (tabPrefix) return tabPrefix;
-
-  if (existsSync(PAGES_CACHE)) {
-    const pages = JSON.parse(readFileSync(PAGES_CACHE, 'utf8'));
-    const existing = pages.find(p => p.url.includes('copilot.microsoft.com'));
-    if (existing) return existing.targetId.slice(0, 8);
-  }
-
+  // Always open a fresh tab to avoid SPA navigation issues
   const list = await cdp(['list']);
-  const firstLine = list.split('\n')[0];
-  if (!firstLine) throw new Error('No Chrome tabs found. Is Chrome running with --remote-debugging-port=9222?');
-  return firstLine.slice(0, 8);
+  const anchor = list.split('\n')[0]?.slice(0, 8);
+  if (!anchor) throw new Error('No Chrome tabs found. Is Chrome running with --remote-debugging-port=9222?');
+  const raw = await cdp(['evalraw', anchor, 'Target.createTarget', '{"url":"about:blank"}']);
+  const { targetId } = JSON.parse(raw);
+  await cdp(['list']); // refresh cache
+  return targetId.slice(0, 8);
 }
 
 async function injectClipboardInterceptor(tab) {
@@ -101,17 +98,11 @@ async function extractAnswer(tab) {
   const answer = await cdp(['eval', tab, `window.__bingClipboard || ''`]);
   if (!answer) throw new Error('Clipboard interceptor returned empty text');
 
-  const raw = await cdp(['eval', tab, `
-    (function() {
-      var sources = Array.from(document.querySelectorAll('${S.sourceLink}'))
-        .map(a => ({ url: a.href, title: a.innerText?.trim().split('\\n')[0] || a.title || '' }))
-        .filter(s => s.url && !s.url.includes('${S.sourceExclude}'))
-        .filter((v, i, arr) => arr.findIndex(x => x.url === v.url) === i)
-        .slice(0, 10);
-      return JSON.stringify(sources);
-    })()
-  `]).catch(() => '[]');
-  const sources = JSON.parse(raw);
+  // Regex parse Markdown links from clipboard — robust against DOM changes
+  const sources = Array.from(answer.matchAll(/\[([^\]]+)\]\((https?:\/\/[^\s\)]+)\)/g))
+    .map(m => ({ title: m[1], url: m[2] }))
+    .filter((v, i, arr) => arr.findIndex(x => x.url === v.url) === i)
+    .slice(0, 10);
 
   return { answer: answer.trim(), sources };
 }

package/extractors/consent.mjs

@@ -102,8 +102,12 @@ const VERIFY_DETECT_JS = `
   
   // --- Generic verification buttons (catch-all) ---
   var btns = Array.from(document.querySelectorAll('button, input[type=submit], a[role=button]'));
-  var verify = btns.find(b => /^(verify|verification|verify you are human|i am human|not a robot|continue|proceed)$/i.test(b.innerText?.trim() || b.value || ''));
-  if (verify && !document.querySelector('iframe[src*="recaptcha"]')) {
+  var verify = btns.find(b => {
+    var t = (b.innerText?.trim() || b.value || '').toLowerCase();
+    return (t.includes('verify') || t.includes('human') || t.includes('robot') || t.includes('continue') || t.includes('proceed')) &&
+           !t.includes('verified') && !document.querySelector('iframe[src*="recaptcha"]');
+  });
+  if (verify) {
     verify.click();
     return 'clicked-verify:' + (verify.innerText?.trim() || verify.value);
   }
@@ -131,7 +135,10 @@ const VERIFY_RETRY_JS = `
   
   // Try clicking any verify/continue button again
   var btns = Array.from(document.querySelectorAll('button, input[type=submit], a[role=button]'));
-  var btn = btns.find(b => /^(verify|continue|next|i am human|not a robot|submit)$/i.test(b.innerText?.trim() || b.value || ''));
+  var btn = btns.find(b => {
+    var t = (b.innerText?.trim() || b.value || '').toLowerCase();
+    return t.includes('verify') || t.includes('human') || t.includes('robot') || t.includes('continue') || t.includes('next') || t.includes('submit');
+  });
   if (btn) { btn.click(); return 'clicked:' + (btn.innerText?.trim() || btn.value); }
   
   // Try Turnstile checkbox

package/extractors/gemini.mjs

@@ -44,13 +44,14 @@ function cdp(args, timeoutMs = 30000) {
 
 async function getOrOpenTab(tabPrefix) {
   if (tabPrefix) return tabPrefix;
-  if (existsSync(PAGES_CACHE)) {
-    const pages = JSON.parse(readFileSync(PAGES_CACHE, 'utf8'));
-    const existing = pages.find(p => p.url.includes('gemini.google.com'));
-    if (existing) return existing.targetId.slice(0, 8);
-  }
+  // Always open a fresh tab to avoid SPA navigation issues
   const list = await cdp(['list']);
-  return list.split('\n')[0].slice(0, 8);
+  const anchor = list.split('\n')[0]?.slice(0, 8);
+  if (!anchor) throw new Error('No Chrome tabs found. Is Chrome running with --remote-debugging-port=9222?');
+  const raw = await cdp(['evalraw', anchor, 'Target.createTarget', '{"url":"about:blank"}']);
+  const { targetId } = JSON.parse(raw);
+  await cdp(['list']); // refresh cache
+  return targetId.slice(0, 8);
 }
 
 async function typeIntoGemini(tab, text) {
@@ -111,53 +112,11 @@ async function extractAnswer(tab) {
   const answer = await cdp(['eval', tab, `window.__geminiClipboard || ''`]);
   if (!answer) throw new Error('Clipboard interceptor returned empty text');
 
-  // Click "Sources" button to open the sidebar with proper source cards
-  const sourceExcludeFilter = S.sourcesExclude.map(e => `!s.url.includes('${e}')`).join(' && ');
-  await cdp(['eval', tab, `
-    (function() {
-      var btn = document.querySelector('${S.sourcesSidebarButton}');
-      if (!btn) btn = Array.from(document.querySelectorAll('button')).find(b => b.innerText?.trim() === 'Sources');
-      if (btn) { btn.click(); return 'clicked'; }
-      return 'not-found';
-    })()
-  `]).catch(() => 'not-found');
-
-  // Wait for the sources sidebar to populate
-  await new Promise(r => setTimeout(r, 1500));
-
-  // Extract sources from the sidebar panel (has proper URLs + titles)
-  const raw = await cdp(['eval', tab, `
-    (function() {
-      // Find the Sources sidebar container by heading
-      var headings = Array.from(document.querySelectorAll('h1, h2, h3, [class*="header"]'));
-      var sourceHeading = headings.find(h => h.innerText?.trim() === 'Sources');
-      if (sourceHeading) {
-        var container = sourceHeading.closest('.container') || sourceHeading.parentElement;
-        var links = Array.from(container.querySelectorAll('a[href^="http"]'))
-          .map(a => ({ url: a.href.split('#')[0], title: a.innerText?.trim().split('\\n')[0] || '' }))
-          .filter(s => s.url && ${sourceExcludeFilter})
-          .filter((v, i, arr) => arr.findIndex(x => x.url === v.url) === i)
-          .slice(0, 8);
-        return JSON.stringify(links);
-      }
-      // Fallback: inline source cards with aria-labels
-      var cards = Array.from(document.querySelectorAll('${S.citationButtonPattern}'));
-      if (cards.length) {
-        return JSON.stringify(cards.map(b => {
-          var label = b.getAttribute('aria-label') || '';
-          var name = label.match(${S.citationNameRegex})?.[1] || label;
-          return { url: '', title: name };
-        }));
-      }
-      // Last resort: page-wide links (may include footer junk)
-      return JSON.stringify(Array.from(document.querySelectorAll('a[href^="http"]'))
-        .map(a => ({ url: a.href.split('#')[0], title: a.innerText?.trim().split('\\n')[0] || '' }))
-        .filter(s => s.url && !s.url.includes('gemini.google') && !s.url.includes('gstatic') && !s.url.includes('google.com/search'))
-        .filter((v, i, arr) => arr.findIndex(x => x.url === v.url) === i)
-        .slice(0, 8));
-    })()
-  `]).catch(() => '[]');
-  const sources = JSON.parse(raw);
+  // Regex parse Markdown links from clipboard — robust against DOM changes
+  const sources = Array.from(answer.matchAll(/\[([^\]]+)\]\((https?:\/\/[^\s\)]+)\)/g))
+    .map(m => ({ title: m[1], url: m[2] }))
+    .filter((v, i, arr) => arr.findIndex(x => x.url === v.url) === i)
+    .slice(0, 10);
 
   return { answer: answer.trim(), sources };
 }