feat(workflow): workflow as tool output schema (#26241)

Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
Co-authored-by: crazywoola <100913391+crazywoola@users.noreply.github.com>
Co-authored-by: Novice <novice12185727@gmail.com>

api/core/tools/entities/tool_bundle.py

@@ -1,4 +1,6 @@
-from pydantic import BaseModel
+from collections.abc import Mapping
+
+from pydantic import BaseModel, Field
 
 from core.tools.entities.tool_entities import ToolParameter
 
@@ -25,3 +27,5 @@ class ApiToolBundle(BaseModel):
     icon: str | None = None
     # openapi operation
     openapi: dict
+    # output schema
+    output_schema: Mapping[str, object] = Field(default_factory=dict)

api/core/tools/utils/workflow_configuration_sync.py

@@ -3,6 +3,7 @@ from typing import Any
 
 from core.app.app_config.entities import VariableEntity
 from core.tools.entities.tool_entities import WorkflowToolParameterConfiguration
+from core.workflow.nodes.base.entities import OutputVariableEntity
 
 
 class WorkflowToolConfigurationUtils:
@@ -24,6 +25,31 @@ class WorkflowToolConfigurationUtils:
 
         return [VariableEntity.model_validate(variable) for variable in start_node.get("data", {}).get("variables", [])]
 
+    @classmethod
+    def get_workflow_graph_output(cls, graph: Mapping[str, Any]) -> Sequence[OutputVariableEntity]:
+        """
+        get workflow graph output
+        """
+        nodes = graph.get("nodes", [])
+        outputs_by_variable: dict[str, OutputVariableEntity] = {}
+        variable_order: list[str] = []
+
+        for node in nodes:
+            if node.get("data", {}).get("type") != "end":
+                continue
+
+            for output in node.get("data", {}).get("outputs", []):
+                entity = OutputVariableEntity.model_validate(output)
+                variable = entity.variable
+
+                if variable not in variable_order:
+                    variable_order.append(variable)
+
+                # Later end nodes override duplicated variable definitions.
+                outputs_by_variable[variable] = entity
+
+        return [outputs_by_variable[variable] for variable in variable_order]
+
     @classmethod
     def check_is_synced(
         cls, variables: list[VariableEntity], tool_configurations: list[WorkflowToolParameterConfiguration]

api/core/tools/workflow_as_tool/provider.py

@@ -162,6 +162,20 @@ class WorkflowToolProviderController(ToolProviderController):
             else:
                 raise ValueError("variable not found")
 
+        # get output schema from workflow
+        outputs = WorkflowToolConfigurationUtils.get_workflow_graph_output(graph)
+
+        reserved_keys = {"json", "text", "files"}
+
+        properties = {}
+        for output in outputs:
+            if output.variable not in reserved_keys:
+                properties[output.variable] = {
+                    "type": output.value_type,
+                    "description": "",
+                }
+        output_schema = {"type": "object", "properties": properties}
+
         return WorkflowTool(
             workflow_as_tool_id=db_provider.id,
             entity=ToolEntity(
@@ -177,6 +191,7 @@ class WorkflowToolProviderController(ToolProviderController):
                     llm=db_provider.description,
                 ),
                 parameters=workflow_tool_parameters,
+                output_schema=output_schema,
             ),
             runtime=ToolRuntime(
                 tenant_id=db_provider.tenant_id,

api/core/tools/workflow_as_tool/tool.py

@@ -114,6 +114,11 @@ class WorkflowTool(Tool):
             for file in files:
                 yield self.create_file_message(file)  # type: ignore
 
+        # traverse `outputs` field and create variable messages
+        for key, value in outputs.items():
+            if key not in {"text", "json", "files"}:
+                yield self.create_variable_message(variable_name=key, variable_value=value)
+
         self._latest_usage = self._derive_usage_from_result(data)
 
         yield self.create_text_message(json.dumps(outputs, ensure_ascii=False))