ThinkSplitter v3: narration gate + paragraph classifier + degenerate-output auto-recovery; num_predict 1100

deploy/draco-mcp.service

@@ -0,0 +1,15 @@
+[Unit]
+Description=DRACO MCP server (streamable-http)
+After=network.target
+
+[Service]
+Type=simple
+WorkingDirectory=/opt/draco
+ExecStart=/opt/draco/venv/bin/python3 -m uvicorn mcp_server:app --host 127.0.0.1 --port 8013 --log-level warning
+Restart=always
+RestartSec=5
+Environment=PYTHONUNBUFFERED=1
+Environment=DRACO_MCP_PORT=8013
+
+[Install]
+WantedBy=multi-user.target

deploy/nginx-draco.conf

@@ -5,6 +5,14 @@ server {
     access_log /var/log/nginx/draco.access.log;
     error_log /var/log/nginx/draco.error.log;
 
+    location = /mcp {
+        proxy_pass http://127.0.0.1:8013;
+        proxy_set_header Host localhost:8013;   # MCP SDK guard requires allowed host[:port]
+        proxy_buffering off;          # SSE
+        proxy_cache off;
+        proxy_read_timeout 300s;
+    }
+
     location / {
         proxy_pass http://127.0.0.1:8012;
         proxy_set_header Host $host;