Merge branch 'feature/mcp-ui' into bot_manager

routes/chat.py

@@ -144,16 +144,13 @@ async def enhanced_generate_stream_response(
                     elif isinstance(msg, ToolMessage) and msg.content:
                         message_tag = "TOOL_RESPONSE"
                         waiting_for_answer_first_char = False
-                        # Always output UIResource, ask_user and render_ui responses even when tool_response is disabled
+                        # Always output UIResource responses even when tool_response is disabled
                         is_ui_resource = (
                             msg.text
                             and msg.text.lstrip().startswith('{"')
                             and '"ui://' in msg.text
-                            and ('"text/html' in msg.text or '"text/uri-list' in msg.text)
                         )
-                        is_ask_user = msg.name == 'ask_user'
-                        is_render_ui = msg.name == 'render_ui'
-                        if config.tool_response or is_ui_resource or is_ask_user or is_render_ui:
+                        if config.tool_response or is_ui_resource:
                             new_content = f"[{message_tag}] {msg.name}\n{msg.text}\n"
 
                     # Collect full content

skills/common/mcp-ui/ui_render_server.py

@@ -8,6 +8,8 @@ import json
 import sys
 from typing import Any, Dict
 
+from mcp_ui_server import create_ui_resource, UIMetadataKey
+
 from mcp_common import (
     create_error_response,
     create_initialize_response,
@@ -18,40 +20,65 @@ from mcp_common import (
 )
 
 
-ASK_USER_RESPONSE = "Questions sent to user."
+def _serialize_ui_resource(ui_resource) -> str:
+    """Serialize a UIResource to JSON string."""
+    return json.dumps(ui_resource.model_dump(mode="json"), ensure_ascii=False)
 
 
 def ask_user() -> Dict[str, Any]:
-    """Return a minimal fixed response for ask_user tool.
+    """Return a UIResource response for ask_user tool.
 
-    The actual questions/options are already in the TOOL_CALL arguments,
-    so the frontend parses them directly from there. This response only
-    serves to acknowledge the tool call and minimize token usage in the
-    subsequent LLM inference round.
+    The actual questions are in the TOOL_CALL arguments. The frontend
+    detects ui_type from the UIResource metadata and extracts content
+    from the corresponding TOOL_CALL args.
     """
-    return {
-        "content": [
-            {"type": "text", "text": ASK_USER_RESPONSE}
-        ]
-    }
+    resource = create_ui_resource({
+        "uri": "ui://mcp-ui/ask-user",
+        "content": {"type": "rawHtml", "htmlString": "Questions sent to user."},
+        "encoding": "text",
+        "uiMetadata": {
+            "type": "ask_user",
+            "interactive": True,
+        },
+    })
+    return {"content": [{"type": "text", "text": _serialize_ui_resource(resource)}]}
 
 
-RENDER_UI_RESPONSE = "UI rendered."
+def render_ui(arguments: Dict[str, Any]) -> Dict[str, Any]:
+    """Return a UIResource response for render_ui tool.
 
-
-def render_ui() -> Dict[str, Any]:
-    """Return a minimal fixed response for render_ui tool.
-
-    The actual html_content/url is already in the TOOL_CALL arguments,
-    so the frontend parses them directly from there. This response only
-    serves to acknowledge the tool call and minimize token usage in the
-    subsequent LLM inference round.
+    The actual html_content/url is in the TOOL_CALL arguments. The frontend
+    detects ui_type from the UIResource metadata and extracts content
+    from the corresponding TOOL_CALL args.
     """
-    return {
-        "content": [
-            {"type": "text", "text": RENDER_UI_RESPONSE}
-        ]
-    }
+    html_content = arguments.get("html_content", "")
+    url = arguments.get("url", "")
+    width = arguments.get("width", "100%")
+    height = arguments.get("height", "auto")
+
+    if html_content:
+        resource = create_ui_resource({
+            "uri": "ui://mcp-ui/render-ui",
+            "content": {"type": "rawHtml", "htmlString": "UI rendered."},
+            "encoding": "text",
+            "uiMetadata": {
+                UIMetadataKey.PREFERRED_FRAME_SIZE: [width, height],
+                "type": "render_ui_html",
+                "interactive": False,
+            },
+        })
+    else:
+        resource = create_ui_resource({
+            "uri": "ui://mcp-ui/render-ui",
+            "content": {"type": "externalUrl", "iframeUrl": url},
+            "encoding": "text",
+            "uiMetadata": {
+                UIMetadataKey.PREFERRED_FRAME_SIZE: [width, height],
+                "type": "render_ui_url",
+                "interactive": False,
+            },
+        })
+    return {"content": [{"type": "text", "text": _serialize_ui_resource(resource)}]}
 
 
 async def handle_request(request: Dict[str, Any]) -> Dict[str, Any]:
@@ -109,7 +136,7 @@ async def handle_request(request: Dict[str, Any]) -> Dict[str, Any]:
                         request_id, -32602, "Missing required parameter: html_content or url"
                     )
 
-                result = render_ui()
+                result = render_ui(arguments)
                 return {"jsonrpc": "2.0", "id": request_id, "result": result}
 
             elif tool_name == "ask_user":