feat: add realtime workflow vision tools

backend/routes/assistants.py

@@ -85,6 +85,7 @@ async def _validate_workflow_references(
     graph = body.graph
     engine = WorkflowEngine(graph)
     settings = graph.get("settings") or {}
+    runtime_mode = str(settings.get("runtimeMode") or "pipeline")
     resource_expectations: dict[str, str] = {}
     vision_resource_ids: set[str] = set()
     for key, capability in (
@@ -118,16 +119,22 @@ async def _validate_workflow_references(
         stage = engine.agent_stage_config(node_id)
         if not stage.vision_enabled:
             continue
-        resource_id = (
-            stage.vision_model_resource_id or stage.llm_resource_id
-        )
+        if stage.vision_model_resource_id:
+            resource_id = stage.vision_model_resource_id
+            capability = "LLM"
+        elif runtime_mode == "realtime":
+            resource_id = str(settings.get("defaultRealtimeResourceId") or "")
+            capability = "Realtime"
+        else:
+            resource_id = stage.llm_resource_id
+            capability = "LLM"
         if not resource_id:
             raise HTTPException(
                 400,
                 f"Agent 节点 {node_id} 开启视觉理解时必须选择"
-                "支持图片输入的大语言模型或视觉模型",
+                "支持图片输入的当前模型或独立视觉模型",
             )
-        resource_expectations[resource_id] = "LLM"
+        resource_expectations[resource_id] = capability
         vision_resource_ids.add(resource_id)
     for resource_id, capability in resource_expectations.items():
         resource = await session.get(ModelResource, resource_id)

backend/routes/model_registry.py

@@ -30,6 +30,19 @@ router = APIRouter(
     tags=["model-registry"],
     dependencies=[Depends(require_admin)],
 )
+NATIVE_IMAGE_UNSUPPORTED_REALTIME_INTERFACES = frozenset(
+    {"qwen-audio-realtime"}
+)
+
+
+def _supports_configurable_image_input(definition: InterfaceDefinition) -> bool:
+    if definition.capability == "LLM":
+        return True
+    return (
+        definition.capability == "Realtime"
+        and definition.interface_type
+        not in NATIVE_IMAGE_UNSUPPORTED_REALTIME_INTERFACES
+    )
 
 
 def _definition_dict(row: InterfaceDefinition) -> dict:
@@ -156,9 +169,11 @@ async def create_model_resource(
         interface_type=definition.interface_type,
         values=body.values,
         secrets=secrets,
-        support_image_input=body.support_image_input
-        if definition.capability == "LLM"
-        else False,
+        support_image_input=(
+            body.support_image_input
+            if _supports_configurable_image_input(definition)
+            else False
+        ),
         enabled=body.enabled,
         is_default=body.is_default,
     )
@@ -244,7 +259,9 @@ async def update_model_resource(
     row.values = body.values
     row.secrets = secrets
     row.support_image_input = (
-        body.support_image_input if definition.capability == "LLM" else False
+        body.support_image_input
+        if _supports_configurable_image_input(definition)
+        else False
     )
     row.enabled = body.enabled
     row.is_default = body.is_default