diff --git a/.agents/references/known-issues.md b/.agents/references/known-issues.md
index d316a31..eae3de9 100644
--- a/.agents/references/known-issues.md
+++ b/.agents/references/known-issues.md
@@ -21,6 +21,15 @@
 > 鍮꾩듂븳 臾몄젣媛 옱諛쒗븯硫 archive뿉꽌 寃깋븯꽭슂.
 
 
+### [2026-04-16] [Extension] ★ AG Native 세션 AI 응답이 Discord에 전혀 전달되지 않음 (미해결, #632)
+- **증상**: Discord에 **명령 승인 신호만 전달**되고, AI의 대화 응답/답변 텍스트는 전혀 전달되지 않음. 수십 세션에 걸쳐 지속 발생.
+- **원인 1 (SDK 경로)**: `GetCascadeTrajectorySteps(cascadeId=세션ID)` → `500 trajectory not found`. AG Native 세션은 Cascade trajectory API에 등록되지 않아 step-probe의 RT-CAPTURE가 불가능. `stepCount`가 항상 1, `delta`는 항상 0이므로 응답 캡처 루프 진입 자체가 안 됨.
+- **원인 2 (DOM 경로)**: Observer `scanChatBodies()`가 `[data-testid="conversation-view"]`, `[data-step-index]`, `.markdown-body`, `.prose` 등을 탐색하지만, **AG Native 렌더러에는 이 셀렉터가 전부 존재하지 않음** (DOM dump 확인: hasConversationView=false, hasStepIndex=false, hasMarkdownBody=false, hasProse=false, dataTestIds=[])
+- **결과**: 버튼 감지(scan→/pending)는 `button` 태그 직접 탐색이므로 정상 작동하나, AI 응답 텍스트 추출 경로는 SDK/DOM 양쪽 모두 구조적으로 차단됨
+- **해결 방향**: AG Native 채팅 패널의 실제 DOM 구조를 deep-inspect로 분석하여, AI 응답 컨테이너의 올바른 셀렉터(class/attribute)를 찾아 `scanChatBodies()` 수정 필요. SDK 경로는 AG 구조적 한계로 사용 불가.
+- **주의**: AG Native 렌더러는 `data-testid`, `data-step-index` 등 Cascade 전용 속성을 사용하지 않음. DOM 분석 시 반드시 AG 패널이 활성화된 상태에서 dump를 취득해야 함 (설정 페이지와 혼동 금지)
+- **Vikunja**: #632
+
 ### [2026-04-16] [Extension] 터미널 출력(stdout) 텍스트가 명령어로 Discord에 전송 (v0.5.50)
 - **증상**: Discord에 `cmd="No extension.log found"`, `cmd="AG CLI not found..."`, `cmd="Log found: C:\..."` 등 터미널 **출력** 텍스트가 명령어로 전송됨
 - **원인**: Observer가 code 블록 2개를 감지: (1) 프롬프트+명령어 → JUNK_CODE_RE로 스킵, (2) 터미널 출력 → 유효한 code로 판단 → description에 포함. http-bridge enrichment에서 description에 prompt marker(`>`)가 없으면 rawDesc 전체를 enrichedCmd로 채택
diff --git a/.agents/workflows/helpers/analyze_dom.py b/.agents/workflows/helpers/analyze_dom.py
new file mode 100644
index 0000000..a27bf61
--- /dev/null
+++ b/.agents/workflows/helpers/analyze_dom.py
@@ -0,0 +1,160 @@
+"""Analyze AG Native DOM structure to find AI response containers."""
+import json, os, sys
+
+def load_dump():
+    bridge = os.path.join(os.path.expanduser('~'), '.gemini', 'antigravity', 'bridge')
+    # Try deep-inspect result first, then dump_html
+    for fname in ['deep-inspect-result.json', 'dump_html.json']:
+        fpath = os.path.join(bridge, fname)
+        if os.path.exists(fpath):
+            print(f"Loading: {fname} ({os.path.getsize(fpath)} bytes)")
+            with open(fpath, 'r', encoding='utf-8-sig') as f:
+                return json.load(f), fname
+    return None, None
+
+def find_text_containers(node, path="", depth=0, results=None):
+    """Recursively find nodes with substantial text content (potential AI response containers)."""
+    if results is None:
+        results = []
+    if not isinstance(node, dict):
+        return results
+    
+    tag = node.get('tag', '')
+    cls = node.get('cls', '')
+    text = node.get('text', '')
+    attrs = node.get('attrs', {})
+    children = node.get('children', [])
+    
+    cur_path = f"{path}/{tag}"
+    if cls:
+        short_cls = cls[:60]
+        cur_path += f".{short_cls}"
+    
+    # Look for nodes with long text (potential AI responses)
+    if text and len(text) > 50:
+        results.append({
+            'path': cur_path,
+            'depth': depth,
+            'tag': tag,
+            'cls': cls[:100],
+            'text_len': len(text),
+            'text_preview': text[:120],
+            'attrs': {k:v for k,v in attrs.items() if k not in ('style',)}
+        })
+    
+    for child in children:
+        find_text_containers(child, cur_path, depth+1, results)
+    
+    return results
+
+def find_by_class_pattern(node, patterns, path="", depth=0, results=None):
+    """Find nodes matching class patterns."""
+    if results is None:
+        results = []
+    if not isinstance(node, dict):
+        return results
+    
+    tag = node.get('tag', '')
+    cls = node.get('cls', '')
+    attrs = node.get('attrs', {})
+    children = node.get('children', [])
+    text = node.get('text', '')
+    
+    cur_path = f"{path}/{tag}" 
+    
+    for pattern in patterns:
+        if pattern.lower() in cls.lower() or pattern.lower() in str(attrs).lower():
+            child_count = len(children)
+            results.append({
+                'path': cur_path,
+                'depth': depth,
+                'tag': tag,
+                'cls': cls[:150],
+                'pattern': pattern,
+                'text_preview': text[:80] if text else '',
+                'child_count': child_count,
+                'attrs': {k:v[:50] for k,v in attrs.items() if k != 'style'}
+            })
+    
+    for child in children:
+        find_by_class_pattern(child, patterns, cur_path, depth+1, results)
+    
+    return results
+
+def analyze_chat_structure(node, path="", depth=0):
+    """Find the chat/conversation area by looking at the main layout."""
+    if not isinstance(node, dict):
+        return
+    tag = node.get('tag', '')
+    cls = node.get('cls', '')
+    children = node.get('children', [])
+    text = node.get('text', '')
+    attrs = node.get('attrs', {})
+    
+    # Print interesting structural nodes at shallow depths
+    if depth <= 6:
+        child_count = len(children)
+        has_text = bool(text and len(text) > 10)
+        info = f"{'  '*depth}{tag}"
+        if cls:
+            info += f" .{cls[:80]}"
+        if attrs:
+            attr_str = ' '.join(f'{k}={v[:30]}' for k,v in attrs.items() if k not in ('style','class'))
+            if attr_str:
+                info += f" [{attr_str}]"
+        info += f" children={child_count}"
+        if has_text:
+            info += f" text=\"{text[:50]}...\""
+        print(info)
+    
+    for child in children:
+        analyze_chat_structure(child, f"{path}/{tag}", depth+1)
+
+data, fname = load_dump()
+if not data:
+    print("No dump file found!")
+    sys.exit(1)
+
+# Handle both dump formats
+body = data.get('body', data)
+qi = data.get('quickInfo', {})
+
+print("=" * 60)
+print("QUICK INFO")
+print("=" * 60)
+if qi:
+    for k, v in qi.items():
+        if k == 'buttons':
+            print(f"buttons ({len(v)}):")
+            for b in v[:15]:
+                print(f"  [{b.get('tag')}] \"{b.get('text','')[:50]}\" visible={b.get('visible')} cls={b.get('cls','')[:60]}")
+        elif k == 'dataAttrs':
+            print(f"dataAttrs: {v[:30]}")
+        else:
+            print(f"{k}: {v}")
+
+print("\n" + "=" * 60)
+print("CHAT-RELATED CLASS PATTERNS")
+print("=" * 60)
+patterns = ['chat', 'message', 'conversation', 'response', 'answer', 'reply',
+            'markdown', 'prose', 'content', 'panel', 'agent', 'assistant',
+            'planner', 'step', 'trajectory', 'bot', 'ai-', 'turn']
+matches = find_by_class_pattern(body, patterns)
+for m in matches:
+    print(f"  [{m['tag']}] cls=\"{m['cls']}\" pattern={m['pattern']} children={m['child_count']} {m.get('attrs',{})}")
+
+print("\n" + "=" * 60)
+print("LONG TEXT NODES (potential AI responses)")  
+print("=" * 60)
+texts = find_text_containers(body)
+texts.sort(key=lambda x: x['text_len'], reverse=True)
+for t in texts[:20]:
+    print(f"  [{t['tag']}] depth={t['depth']} len={t['text_len']} cls=\"{t['cls'][:60]}\"")
+    print(f"    text: \"{t['text_preview']}\"")
+    if t['attrs']:
+        print(f"    attrs: {t['attrs']}")
+
+print("\n" + "=" * 60)
+print("DOM TREE (depth<=6)")
+print("=" * 60)
+analyze_chat_structure(body)
diff --git a/.agents/workflows/helpers/parse_dump.py b/.agents/workflows/helpers/parse_dump.py
new file mode 100644
index 0000000..68c9d41
--- /dev/null
+++ b/.agents/workflows/helpers/parse_dump.py
@@ -0,0 +1,19 @@
+import json, os, sys
+
+dump_path = os.path.join(os.path.expanduser('~'), '.gemini', 'antigravity', 'bridge', 'dump_html.json')
+with open(dump_path, 'r', encoding='utf-8') as f:
+    data = json.load(f)
+
+qi = data.get('quickInfo', {})
+print('=== Quick Info ===')
+print('hasConversationView:', qi.get('hasConversationView'))
+print('hasStepIndex:', qi.get('hasStepIndex'))
+print('hasBotColor:', qi.get('hasBotColor'))
+print('hasMarkdownBody:', qi.get('hasMarkdownBody'))
+print('hasProse:', qi.get('hasProse'))
+print('totalElements:', qi.get('totalElements'))
+print('dataTestIds:', qi.get('dataTestIds'))
+print('dataAttrs (first 20):', qi.get('dataAttrs', [])[:20])
+print('buttons (first 10):')
+for b in qi.get('buttons', [])[:10]:
+    print(f"  [{b.get('tag')}] {b.get('text', '')[:60]}  visible={b.get('visible')}")
diff --git a/.agents/workflows/helpers/search_dom.py b/.agents/workflows/helpers/search_dom.py
new file mode 100644
index 0000000..bf0537a
--- /dev/null
+++ b/.agents/workflows/helpers/search_dom.py
@@ -0,0 +1,83 @@
+"""Search AG Native DOM dump for chat content and buttons."""
+import json, os
+
+fpath = os.path.join(os.path.expanduser('~'), '.gemini', 'antigravity', 'bridge', 'dump_html_5.json')
+with open(fpath, 'r', encoding='utf-8-sig') as f:
+    data = json.load(f)
+
+body = data.get('body', data.get('bodyTree', {}))
+qi = data.get('quickInfo', {})
+
+# Show all buttons
+print('=== BUTTONS ===')
+for b in qi.get('buttons', []):
+    print(f'  [{b["tag"]}] "{b["text"][:60]}" visible={b["visible"]} cls={b.get("cls","")[:80]}')
+
+# Data attrs
+print('\n=== DATA ATTRS ===')
+for attr in qi.get('dataAttrs', []):
+    print(f'  {attr}')
+
+# Recursive search for nodes by text
+def find_nodes_by_text(node, target, path='', results=None, depth=0):
+    if results is None: results = []
+    if not isinstance(node, dict): return results
+    tag = node.get('tag','')
+    cls = node.get('cls','')
+    text = node.get('text','')
+    children = node.get('children', [])
+    cur = f'{path}/{tag}'
+    if target.lower() in text.lower():
+        results.append({'path': cur, 'depth': depth, 'cls': cls[:80], 'text': text[:80], 'children': len(children)})
+    for c in children:
+        find_nodes_by_text(c, target, cur, results, depth+1)
+    return results
+
+print('\n=== NODES containing "Always run" ===')
+matches = find_nodes_by_text(body, 'Always run')
+for m in matches:
+    print(f'  depth={m["depth"]} cls="{m["cls"]}" text="{m["text"]}" children={m["children"]}')
+
+print('\n=== NODES containing "Always" ===')
+matches = find_nodes_by_text(body, 'Always')
+for m in matches:
+    print(f'  depth={m["depth"]} cls="{m["cls"]}" text="{m["text"]}" children={m["children"]}')
+
+# Find ALL text nodes with > 30 chars
+def find_all_text(node, results=None, depth=0, path=''):
+    if results is None: results = []
+    if not isinstance(node, dict): return results
+    tag = node.get('tag','')
+    cls = node.get('cls','')
+    text = node.get('text','')
+    children = node.get('children', [])
+    if text and len(text) > 30:
+        results.append({'depth': depth, 'tag': tag, 'cls': cls[:80], 'text': text[:100], 'path': f'{path}/{tag}'})
+    for c in children:
+        find_all_text(c, results, depth+1, f'{path}/{tag}')
+    return results
+
+print('\n=== LONG TEXT NODES (>30 chars) ===')
+texts = find_all_text(body)
+texts.sort(key=lambda x: len(x['text']), reverse=True)
+for t in texts[:25]:
+    print(f'  d={t["depth"]} [{t["tag"]}] cls="{t["cls"][:50]}" len={len(t["text"])} "{t["text"][:80]}"')
+
+# Find nodes with many children (structural containers)
+def find_containers(node, results=None, depth=0, path=''):
+    if results is None: results = []
+    if not isinstance(node, dict): return results
+    tag = node.get('tag','')
+    cls = node.get('cls','')
+    children = node.get('children', [])
+    if len(children) > 5:
+        results.append({'depth': depth, 'tag': tag, 'cls': cls[:100], 'children': len(children), 'path': f'{path}/{tag}'})
+    for c in children:
+        find_containers(c, results, depth+1, f'{path}/{tag}')
+    return results
+
+print('\n=== CONTAINERS (>5 children) ===')
+conts = find_containers(body)
+conts.sort(key=lambda x: x['children'], reverse=True)
+for c in conts[:20]:
+    print(f'  d={c["depth"]} [{c["tag"]}] children={c["children"]} cls="{c["cls"][:70]}"')
diff --git a/.agents/workflows/helpers/trace_dom.py b/.agents/workflows/helpers/trace_dom.py
new file mode 100644
index 0000000..bd09074
--- /dev/null
+++ b/.agents/workflows/helpers/trace_dom.py
@@ -0,0 +1,109 @@
+"""Trace the DOM path from body to AI response container."""
+import json, os
+
+fpath = os.path.join(os.path.expanduser('~'), '.gemini', 'antigravity', 'bridge', 'dump_html_5.json')
+with open(fpath, 'r', encoding='utf-8-sig') as f:
+    data = json.load(f)
+
+body = data.get('body', data.get('bodyTree', {}))
+
+def find_path_to_class(node, target_cls, path=None, depth=0):
+    """Find the DOM path down to a node with a matching class."""
+    if path is None: path = []
+    if not isinstance(node, dict): return []
+    
+    tag = node.get('tag', '')
+    cls = node.get('cls', '')
+    children = node.get('children', [])
+    text = node.get('text', '')
+    attrs = node.get('attrs', {})
+    
+    entry = {
+        'depth': depth,
+        'tag': tag,
+        'cls': cls[:120],
+        'children': len(children),
+        'text': text[:60] if text else '',
+        'attrs': {k:v[:40] for k,v in attrs.items() if k not in ('style',)}
+    }
+    
+    if target_cls.lower() in cls.lower():
+        return path + [entry]
+    
+    for i, child in enumerate(children):
+        result = find_path_to_class(child, target_cls, path + [entry], depth+1)
+        if result:
+            return result
+    
+    return []
+
+# Find path to the AI response container
+print("=== PATH TO 'leading-relaxed select-text' ===")
+path = find_path_to_class(body, 'leading-relaxed select-text')
+for p in path:
+    indent = '  ' * p['depth']
+    print(f'{indent}[{p["tag"]}] cls="{p["cls"]}" children={p["children"]} {p["attrs"]}')
+    if p['text']:
+        print(f'{indent}  text: "{p["text"]}"')
+
+# Now get the full subtree of the AI response container
+def get_subtree(node, target_cls, depth=0):
+    if not isinstance(node, dict): return None
+    cls = node.get('cls', '')
+    if target_cls.lower() in cls.lower():
+        return node
+    for child in node.get('children', []):
+        result = get_subtree(child, target_cls, depth+1)
+        if result:
+            return result
+    return None
+
+print("\n=== AI RESPONSE CONTAINER SUBTREE ===")
+container = get_subtree(body, 'leading-relaxed select-text')
+if container:
+    def print_tree(node, depth=0, max_depth=4):
+        if not isinstance(node, dict) or depth > max_depth: return
+        tag = node.get('tag','')
+        cls = node.get('cls','')[:80]
+        text = node.get('text','')
+        children = node.get('children', [])
+        indent = '  ' * depth
+        line = f'{indent}[{tag}]'
+        if cls: line += f' cls="{cls}"'
+        line += f' children={len(children)}'
+        if text: line += f' text="{text[:60]}"'
+        print(line)
+        for c in children:
+            print_tree(c, depth+1, max_depth)
+    
+    print_tree(container, 0, 3)
+
+# Also search for the chat panel container - what wraps the entire conversation
+print("\n=== SEARCH FOR CHAT PANEL WRAPPERS ===")
+chat_patterns = ['chat', 'antigravity', 'gemini', 'panel', 'agentview', 'sidebar', 'conversation']
+for pat in chat_patterns:
+    path = find_path_to_class(body, pat)
+    if path:
+        last = path[-1]
+        print(f'  Pattern "{pat}" found at depth={last["depth"]} [{last["tag"]}] cls="{last["cls"]}" children={last["children"]}')
+
+# Find the parent chain from body to the container - look by scanning ALL class names
+print("\n=== ALL UNIQUE CLASS NAMES (depth <= 12) ===")
+all_classes = set()
+def collect_classes(node, depth=0, max_depth=12):
+    if not isinstance(node, dict) or depth > max_depth: return
+    cls = node.get('cls', '')
+    if cls:
+        for c in cls.split():
+            if len(c) > 3 and not c.startswith('{') and 'mtk' not in c:
+                all_classes.add(c)
+    for child in node.get('children', []):
+        collect_classes(child, depth+1, max_depth)
+
+collect_classes(body)
+# Print classes sorted, grouped by potential relevance
+relevant = sorted([c for c in all_classes if any(k in c.lower() for k in 
+    ['chat', 'message', 'response', 'agent', 'gemini', 'turn', 'model', 'user', 'bot', 'conversation', 'markdown', 'prose', 'text-', 'content'])])
+print("Relevant classes:")
+for c in relevant:
+    print(f'  {c}')
diff --git a/extension/package.json b/extension/package.json
index 8b4bbc6..020809f 100644
--- a/extension/package.json
+++ b/extension/package.json
@@ -2,7 +2,7 @@
   "name": "gravity-bridge",
   "displayName": "Gravity Bridge",
   "description": "Discord-based unified approval system for Antigravity AI interactions.",
-  "version": "0.5.50",
+  "version": "0.5.51",
   "publisher": "variet",
   "engines": {
     "vscode": "^1.100.0"
diff --git a/extension/src/observer-script.ts b/extension/src/observer-script.ts
index 463595d..a6c4ee7 100644
--- a/extension/src/observer-script.ts
+++ b/extension/src/observer-script.ts
@@ -1,7 +1,7 @@
 export function generateApprovalObserverScript(_port: number): string {
   return `
-// ── Gravity Bridge v14: Strict Scope + Junk Filter ──
-// v14: Strict 5-level DOM scope, CSS/source code/icon-glue filters, no fallback
+// ── Gravity Bridge v15: AG Native Chat Relay ──
+// v15: AG Native #conversation + .leading-relaxed.select-text chat body scanning
 (function(){
   'use strict';
   var BASE='',_obs=false,_sent={},_ready=false;
@@ -10,7 +10,7 @@ export function generateApprovalObserverScript(_port: number): string {
   var CLEANUP_MS=300000;
 
   function log(m){console.log('[GB Observer] '+m);}
-  log('v14 Script loaded — Strict Scope + Junk Filter');
+  log('v15 Script loaded — AG Native Chat Relay');
 
   // DIAGNOSTIC BEACON: immediate POST to confirm script execution in renderer
   try {
@@ -460,15 +460,16 @@ export function generateApprovalObserverScript(_port: number): string {
   }
 
   // ══════════════════════════════════════════════════════════════════
-  // v7: STEP-AWARE CHAT BODY SCANNING
-  // Scans [data-step-index] elements inside [data-testid="conversation-view"]
-  // Extracts AI response text while filtering UI noise
+  // v15: AG-NATIVE + CASCADE DUAL CHAT BODY SCANNING
+  // AG Native: #conversation > ... > .leading-relaxed.select-text
+  // Cascade:   [data-testid="conversation-view"] > [data-step-index]
   // ══════════════════════════════════════════════════════════════════
 
   var _lastScrapedStepIndex = -1;
   var _lastStepText = '';
   var _lastStepTextTime = 0;
   var _lastStepTextSent = false;
+  var _lastResponseBlockCount = 0; // track number of response blocks for AG Native
 
   function extractCleanStepText(stepEl) {
     if (!stepEl) return '';
@@ -495,7 +496,7 @@ export function generateApprovalObserverScript(_port: number): string {
     }
 
     // Try to get text from markdown rendering area first
-    // Look for known markdown container patterns
+    // AG Native uses .leading-relaxed.select-text, Cascade uses .markdown-body/.prose
     var mdEl = clone.querySelector('.markdown-body, .prose, [class*="markdown"], [class*="rendered"]');
     var rawText = '';
     if (mdEl && mdEl.innerText && mdEl.innerText.trim().length > 10) {
@@ -515,8 +516,80 @@ export function generateApprovalObserverScript(_port: number): string {
     // One-time DOM dump
     dumpDOMStructure();
 
-    // PRIMARY: Find conversation-view container
-    var cv = document.querySelector('[data-testid="conversation-view"]');
+    // ── STRATEGY 1: AG Native — #conversation or .antigravity-agent-side-panel ──
+    var cv = document.querySelector('#conversation');
+    if (!cv) {
+      cv = document.querySelector('.antigravity-agent-side-panel');
+    }
+
+    if (cv) {
+      // AG Native path: find AI response blocks by class pattern
+      // DOM structure: #conversation > ... > .leading-relaxed.select-text (AI response text)
+      var responseBlocks = cv.querySelectorAll('.leading-relaxed.select-text');
+
+      if (responseBlocks.length > 0) {
+        // Process the LAST (most recent) response block
+        var lastBlock = responseBlocks[responseBlocks.length - 1];
+
+        // Skip if already scraped
+        if (lastBlock.dataset.agChatScraped === 'true' || lastBlock.dataset.agChatScraped === 'pending') {
+          // Check for NEW blocks since last scrape
+          if (responseBlocks.length > _lastResponseBlockCount) {
+            // New block appeared — process it
+            for (var rbi = responseBlocks.length - 1; rbi >= 0; rbi--) {
+              if (responseBlocks[rbi].dataset.agChatScraped !== 'true' && responseBlocks[rbi].dataset.agChatScraped !== 'pending') {
+                lastBlock = responseBlocks[rbi];
+                break;
+              }
+            }
+            if (lastBlock.dataset.agChatScraped === 'true' || lastBlock.dataset.agChatScraped === 'pending') return;
+          } else {
+            return; // Already scraped, no new blocks
+          }
+        }
+
+        var blockText = extractCleanStepText(lastBlock);
+        if (blockText && blockText.length > 30) {
+          // QUALITY CHECK: Skip if the text is mostly short lines (UI artifacts)
+          var lines = blockText.split('\\n').filter(function(l) { return l.trim().length > 0; });
+          var longLines = lines.filter(function(l) { return l.trim().length > 20; });
+          if (longLines.length === 0) {
+            log('AG-Native: skipped (no long lines, likely UI noise)');
+            return;
+          }
+
+          // Wait for content to stabilize (3s no change)
+          if (_lastStepText !== blockText) {
+            _lastStepText = blockText;
+            _lastStepTextTime = Date.now();
+            _lastStepTextSent = false;
+            return; // Wait for next scan cycle
+          }
+
+          if (_lastStepTextSent) return;
+          if (Date.now() - _lastStepTextTime < 3000) return; // Still waiting
+
+          // Content is stable — send it
+          _lastStepTextSent = true;
+          _lastResponseBlockCount = responseBlocks.length;
+          lastBlock.dataset.agChatScraped = 'pending';
+
+          log('AG-Native chat relay: blocks=' + responseBlocks.length + ' text=' + blockText.length + ' chars');
+          (function(el, txt, count) {
+            fetch(BASE + '/chat', {
+              method: 'POST',
+              headers: { 'Content-Type': 'application/json' },
+              body: JSON.stringify({ text: txt, source: 'ag_native_block_' + count, block_index: count })
+            }).then(function() { el.dataset.agChatScraped = 'true'; log('AG-Native chat sent OK'); })
+              .catch(function(e) { el.dataset.agChatScraped = 'false'; log('AG-Native chat send error: ' + e.message); });
+          })(lastBlock, blockText, responseBlocks.length);
+        }
+        return; // AG Native path handled — don't fall through to Cascade path
+      }
+    }
+
+    // ── STRATEGY 2: Cascade — [data-testid="conversation-view"] ──
+    cv = document.querySelector('[data-testid="conversation-view"]');
     if (!cv) {
       // FALLBACK: Try older selectors
       cv = document.querySelector('[class*="conversation"], [class*="chat-container"]');