diff --git a/bin/agent-health.sh b/bin/agent-health.sh
index 00c877a..97d74a4 100755
--- a/bin/agent-health.sh
+++ b/bin/agent-health.sh
@@ -1,4 +1,8 @@
 #!/bin/bash
+# GOLDEN PATH: container -> VM (34.139.37.135) -> bl (100.123.153.75) -> netns -> browser -> agent
+# This script is the operator's heartbeat. If it stops, agents go dark.
+# When debugging: trace each hop. Don't assume -- verify.
+
 # Operator health monitor for Muse agents on bl.
 # Checks each agent via API every 5 minutes. If unresponsive:
 # 1. Restart the browser
diff --git a/bin/muse-chat-api.py b/bin/muse-chat-api.py
index b93abcf..1a64a39 100755
--- a/bin/muse-chat-api.py
+++ b/bin/muse-chat-api.py
@@ -1,5 +1,7 @@
 #!/usr/bin/env python3
-"""
+# GOLDEN PATH: container -> VM (34.139.37.135) -> bl (100.123.153.75) -> netns -> browser -> agent
+# This API is the bridge. Every call traverses 4 hops. Respect the path.
+
 Multi-account muse.ai chat API with approval handling.
 
 Approvals: The browser may show permission dialogs (e.g., "Allow pip to share
@@ -18,6 +20,7 @@ import json, urllib.request, websocket, time, sys, argparse
 ACCOUNTS = {
     "muse": ("muse", "http://127.0.0.1:9410/json/list"),
     "pip": ("pip", "http://127.0.0.1:9420/json/list"),
+    "646": ("646", "http://10.201.202.2:9353/json/list"),
 }
 
 # IPs we trust for auto-approval (our infrastructure)