feat(workflow): add per-agent vision configuration

backend/routes/assistants.py

@@ -16,6 +16,7 @@ from schemas import AssistantOut, AssistantUpsert
 from services.auth import require_admin
 from services.masking import mask, resolve_incoming_key
 from services.node_specs import graph_references, normalize_graph, validate_graph
+from services.workflow_engine import WorkflowEngine
 from sqlalchemy import select
 from sqlalchemy.ext.asyncio import AsyncSession
 
@@ -35,6 +36,10 @@ def _validate_workflow(body: AssistantUpsert) -> None:
     errors = validate_graph(body.graph)
     if errors:
         raise HTTPException(400, "工作流校验失败:" + ";".join(errors))
+    # Graph settings are the source of truth. The flat flag is only a session
+    # capability summary used by clients to decide whether to open the camera.
+    body.vision_enabled = WorkflowEngine(body.graph).uses_vision()
+    body.vision_model_resource_id = None
     refs = graph_references(body.graph)
     body.tool_ids = list(dict.fromkeys([*body.tool_ids, *sorted(refs["tools"])]))
 
@@ -45,8 +50,10 @@ async def _validate_workflow_references(
     if body.type != "workflow" or not body.graph.get("nodes"):
         return
     graph = body.graph
+    engine = WorkflowEngine(graph)
     settings = graph.get("settings") or {}
     resource_expectations: dict[str, str] = {}
+    vision_resource_ids: set[str] = set()
     for key, capability in (
         ("defaultLlmResourceId", "LLM"),
         ("defaultAsrResourceId", "ASR"),
@@ -71,10 +78,32 @@ async def _validate_workflow_references(
             resource_expectations[str(data["ttsResourceId"])] = "TTS"
         if data.get("knowledgeBaseId"):
             knowledge_ids.add(str(data["knowledgeBaseId"]))
+    for node_id, node in engine.nodes.items():
+        if node.get("type") != "agent":
+            continue
+        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 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",
+            )
+        resource_expectations[resource_id] = "LLM"
+        vision_resource_ids.add(resource_id)
     for resource_id, capability in resource_expectations.items():
         resource = await session.get(ModelResource, resource_id)
         if not resource or not resource.enabled or resource.capability != capability:
             raise HTTPException(400, f"Workflow 引用了无效的 {capability} 资源:{resource_id}")
+        if resource_id in vision_resource_ids and not resource.support_image_input:
+            raise HTTPException(
+                400,
+                f"Workflow 视觉模型必须支持图片输入:{resource_id}",
+            )
     for knowledge_id in knowledge_ids:
         knowledge = await session.get(KnowledgeBase, knowledge_id)
         if not knowledge or knowledge.status != "active":
@@ -84,6 +113,9 @@ async def _validate_workflow_references(
 async def _validate_vision_model(
     session: AsyncSession, body: AssistantUpsert
 ) -> None:
+    # Workflow 的视觉模型属于图中的全局/Agent 配置，由上面的图引用校验处理。
+    if body.type == "workflow":
+        return
     if body.vision_enabled:
         if body.vision_model_resource_id:
             resource = await session.get(ModelResource, body.vision_model_resource_id)

backend/routes/voice_webrtc.py

@@ -163,8 +163,13 @@ async def _handle_offer_payload(payload, peers):
     else:
         cfg = await _resolve_config(offer)  # 解析放在建连前,配置错就别建连
         # 服务端助手配置是视觉理解的唯一授权来源；客户端 offer 只负责携带媒体轨。
-        vision_enabled = cfg.vision_enabled
-        if vision_enabled:
+        if cfg.type == "workflow":
+            from services.workflow_engine import WorkflowEngine
+
+            vision_enabled = WorkflowEngine(cfg.graph).uses_vision()
+        else:
+            vision_enabled = cfg.vision_enabled
+        if vision_enabled and cfg.type != "workflow":
             has_native_vision = (
                 not cfg.vision_model_resource_id and cfg.llm_support_image_input
             )