feat: add mcp tool type

feat: add mcp tool type

Author: Nov1c444

python/dify_plugin/entities/tool.py

@@ -394,6 +394,7 @@ class ToolProviderType(Enum):
     API = "api"
     APP = "app"
     DATASET_RETRIEVAL = "dataset-retrieval"
+    MCP = "mcp"
 
     @classmethod
     def value_of(cls, value: str) -> "ToolProviderType":

python/examples/agent/README.md

@@ -21,4 +21,4 @@ ReAct alternates between the LLM reasoning about the situation and taking action
 #### Pros:
 - **Leverages External Information:** Effectively uses external tools to gather information for tasks the model cannot handle alone.
 - **Explainable Reasoning:** Interwoven reasoning and action steps allow some tracking of the Agent's process.
-- **Wide Applicability:** Suitable for tasks requiring external knowledge or specific actions, such as Q&A, information retrieval, and task execution.
+- **Wide Applicability:** Suitable for tasks requiring external knowledge or specific actions, such as Q&A, information retrieval, and task execution.

python/examples/agent/manifest.yaml

@@ -1,11 +1,11 @@
-version: 0.0.10
+version: 0.0.15
 type: plugin
 author: "langgenius"
 name: "agent"
 label:
   en_US: "Dify Agent Strategies"
   zh_Hans: "Dify Agent 策略"
-created_at: "2024-07-12T08:03:44.658609186Z"
+created_at: "2025-01-08T15:22:00.000000000Z"
 icon: icon.svg
 description:
   en_US: Dify official Agent strategies collection

python/examples/agent/requirements.txt

@@ -1 +1 @@
-dify_plugin==0.0.1b65
+dify_plugin==0.2.1

python/examples/agent/strategies/ReAct.py

@@ -50,8 +50,8 @@ class AgentPromptEntity(BaseModel):
 
 
 class ReActAgentStrategy(AgentStrategy):
-    def __init__(self, session):
-        super().__init__(session)
+    def __init__(self, session, runtime):
+        super().__init__(session, runtime)
         self.query = ""
         self.instruction = ""
         self.history_prompt_messages = []

python/examples/agent/strategies/function_calling.py

@@ -23,7 +23,12 @@
     UserPromptMessage,
 )
 from dify_plugin.entities.tool import LogMetadata, ToolInvokeMessage, ToolProviderType
-from dify_plugin.interfaces.agent import AgentModelConfig, AgentStrategy, ToolEntity, ToolInvokeMeta
+from dify_plugin.interfaces.agent import (
+    AgentModelConfig,
+    AgentStrategy,
+    ToolEntity,
+    ToolInvokeMeta,
+)
 
 
 class FunctionCallingParams(BaseModel):
@@ -35,8 +40,8 @@ class FunctionCallingParams(BaseModel):
 
 
 class FunctionCallingAgentStrategy(AgentStrategy):
-    def __init__(self, session):
-        super().__init__(session)
+    def __init__(self, runtime, session):
+        super().__init__(runtime, session)
         self.query = ""
         self.instruction = ""
 
@@ -59,6 +64,8 @@ def _invoke(self, parameters: dict[str, Any]) -> Generator[AgentInvokeMessage]:
         self.query = query
         self.instruction = fc_params.instruction
         history_prompt_messages = fc_params.model.history_prompt_messages
+        history_prompt_messages.insert(0, self._system_prompt_message)
+        history_prompt_messages.append(self._user_prompt_message)
 
         # convert tool messages
         tools = fc_params.tools
@@ -96,7 +103,8 @@ def _invoke(self, parameters: dict[str, Any]) -> Generator[AgentInvokeMessage]:
             )
             yield round_log
 
-            if iteration_step == max_iteration_steps:
+            # If max_iteration_steps=1, need to execute tool calls
+            if iteration_step == max_iteration_steps and max_iteration_steps > 1:
                 # the last iteration, remove all tools
                 prompt_messages_tools = []
 
@@ -195,7 +203,7 @@ def _invoke(self, parameters: dict[str, Any]) -> Generator[AgentInvokeMessage]:
                 data={
                     "output": response,
                     "tool_name": tool_call_names,
-                    "tool_input": {tool_call[1]: tool_call[2] for tool_call in tool_calls},
+                    "tool_input": [{"name": tool_call[1], "args": tool_call[2]} for tool_call in tool_calls],
                 },
                 metadata={
                     LogMetadata.STARTED_AT: model_started_at,
@@ -208,28 +216,30 @@ def _invoke(self, parameters: dict[str, Any]) -> Generator[AgentInvokeMessage]:
                 },
             )
             assistant_message = AssistantPromptMessage(content="", tool_calls=[])
-            if tool_calls:
-                assistant_message.tool_calls = [
-                    AssistantPromptMessage.ToolCall(
-                        id=tool_call[0],
-                        type="function",
-                        function=AssistantPromptMessage.ToolCall.ToolCallFunction(
-                            name=tool_call[1],
-                            arguments=json.dumps(tool_call[2], ensure_ascii=False),
-                        ),
-                    )
-                    for tool_call in tool_calls
-                ]
-            else:
+            if not tool_calls:
                 assistant_message.content = response
-
-            current_thoughts.append(assistant_message)
+                current_thoughts.append(assistant_message)
 
             final_answer += response + "\n"
 
             # call tools
             tool_responses = []
             for tool_call_id, tool_call_name, tool_call_args in tool_calls:
+                current_thoughts.append(
+                    AssistantPromptMessage(
+                        content="",
+                        tool_calls=[
+                            AssistantPromptMessage.ToolCall(
+                                id=tool_call_id,
+                                type="function",
+                                function=AssistantPromptMessage.ToolCall.ToolCallFunction(
+                                    name=tool_call_name,
+                                    arguments=json.dumps(tool_call_args, ensure_ascii=False),
+                                ),
+                            )
+                        ],
+                    )
+                )
                 tool_instance = tool_instances[tool_call_name]
                 tool_call_started_at = time.perf_counter()
                 tool_call_log = self.create_log_message(
@@ -341,8 +351,11 @@ def _invoke(self, parameters: dict[str, Any]) -> Generator[AgentInvokeMessage]:
                     LogMetadata.TOTAL_TOKENS: current_llm_usage.total_tokens if current_llm_usage else 0,
                 },
             )
+            # If max_iteration_steps=1, need to return tool responses
+            if tool_responses and max_iteration_steps == 1:
+                for resp in tool_responses:
+                    yield self.create_text_message(resp["tool_response"])
             iteration_step += 1
-
         yield self.create_json_message(
             {
                 "execution_metadata": {
@@ -452,8 +465,6 @@ def _organize_prompt_messages(
         current_thoughts: list[PromptMessage],
         history_prompt_messages: list[PromptMessage],
     ) -> list[PromptMessage]:
-        history_prompt_messages.insert(0, self._system_prompt_message)
-        history_prompt_messages.append(self._user_prompt_message)
         prompt_messages = [
             *history_prompt_messages,
             *current_thoughts,