fix(agent): optimize of base agent

dbgpt/agent/core/action/base.py

@@ -53,6 +53,8 @@ class ActionOutput(BaseModel):
     ask_user: Optional[bool] = False
     # 如果当前agent能确定下个发言者，需要在这里指定
     next_speakers: Optional[List[str]] = None
+    # 强制重试，不受重试次数影响限制
+    force_retry: Optional[bool] = False
 
     @model_validator(mode="before")
     @classmethod

dbgpt/agent/core/base_agent.py

@@ -47,6 +47,7 @@ class ConversableAgent(Role, Agent):
     stream_out: bool = True
     # 确认当前Agent是否需要进行参考资源展示
     show_reference: bool = False
+    name_prefix: Optional[str] = None
 
     executor: Executor = Field(
         default_factory=lambda: ThreadPoolExecutor(max_workers=1),
@@ -58,6 +59,13 @@ def __init__(self, **kwargs):
         Role.__init__(self, **kwargs)
         Agent.__init__(self)
 
+    @property
+    def name(self) -> str:
+        """Return the name of the agent."""
+        if self.name_prefix is not None:
+            return f"{self.current_profile.get_name()}[{self.name_prefix}]"
+        return self.current_profile.get_name()
+
     def check_available(self) -> None:
         """Check if the agent is available.
 
@@ -347,7 +355,22 @@ async def generate_reply(
             fail_reason = None
             current_retry_counter = 0
             is_success = True
-            while current_retry_counter < self.max_retry_count:
+            force_retry = False
+            while force_retry or current_retry_counter < self.max_retry_count:
+                # Action force_retry 强制重试的处理: Action中明确指定需要进行重试，重试的消息按以下规则重新生成
+                # - 重新生成消息，保留上一轮的Action，并增加Rounds
+                # - 将上一轮的Action的Content作为当前输入消息
+                if force_retry:
+                    if reply_message.action_report is None:
+                        raise ValueError("action output is None when force_retry")
+                    received_message.content = reply_message.action_report.content
+                    received_message.rounds = reply_message.rounds + 1
+                    reply_message = self._init_reply_message(
+                        received_message=received_message,
+                        rely_messages=rely_messages,
+                    )
+
+                # 普通重试的处理
                 if current_retry_counter > 0:
                     retry_message = self._init_reply_message(
                         received_message=received_message,
@@ -460,6 +483,21 @@ async def generate_reply(
 
                 question: str = received_message.content or ""
                 ai_message: str = llm_reply or ""
+
+                # force_retry means this reply do not complete
+                # should reentry and do more things
+                force_retry = False
+                if act_out is not None and act_out.force_retry:
+                    await self.write_memories(
+                        question=question,
+                        ai_message=ai_message,
+                        action_output=act_out,
+                        check_pass=check_pass,
+                        check_fail_reason=fail_reason,
+                    )
+                    force_retry = True
+                    continue
+
                 # 5.Optimize wrong answers myself
                 if not check_pass:
                     if not act_out.have_retry:

dbgpt/agent/core/memory/gpts/default_gpts_memory.py

@@ -146,4 +146,7 @@ def get_by_conv_id(self, conv_id: str) -> List[GptsMessage]:
 
     def get_last_message(self, conv_id: str) -> Optional[GptsMessage]:
         """Get the last message in the conversation."""
-        return None
+        messages: List[GptsMessage] = self.get_by_conv_id(conv_id)
+        if messages is None or len(messages) == 0:
+            return None
+        return messages[-1]

dbgpt/agent/core/role.py

@@ -233,12 +233,13 @@ async def write_memories(
             raise ValueError("Action output is required to save to memory.")
 
         mem_thoughts = action_output.thoughts or ai_message
-        observation = action_output.observations
+        observation = check_fail_reason or action_output.observations
+        action = action_output.action
 
         memory_map = {
             "question": question,
             "thought": mem_thoughts,
-            "action": check_fail_reason,
+            "action": action,
             "observation": observation,
         }
         write_memory_template = self.write_memory_template

dbgpt/agent/expand/actions/tool_action.py

@@ -21,12 +21,14 @@ class ToolInput(BaseModel):
     tool_name: str = Field(
         ...,
         description="The name of a tool that can be used to answer the current question"
-        " or solve the current task.",
+        " or solve the current task. "
+        "If no suitable tool is selected, leave this blank.",
     )
     args: dict = Field(
         default={"arg name1": "", "arg name2": ""},
         description="The tool selected for the current target, the parameter "
-        "information required for execution",
+        "information required for execution, "
+        "If no suitable tool is selected, leave this blank.",
     )
     thought: str = Field(..., description="Summary of thoughts to the user")
 
@@ -68,9 +70,9 @@ def ai_out_schema(self) -> Optional[str]:
         }
 
         return f"""Please response in the following json format:
-        {json.dumps(out_put_schema, indent=2, ensure_ascii=False)}
-        Make sure the response is correct json and can be parsed by Python json.loads.
-        """
+{json.dumps(out_put_schema, indent=2, ensure_ascii=False)}
+Make sure the response is correct json and can be parsed by Python json.loads.
+and do not write the comment in json，only write the json content."""
 
     async def run(
         self,
@@ -91,6 +93,15 @@ async def run(
             need_vis_render (bool, optional): Whether need visualization rendering.
                 Defaults to True.
         """
+        success, error = parse_json_safe(ai_message)
+        if not success:
+            return ActionOutput(
+                is_exe_success=False,
+                content=f"Tool Action execute failed! llm reply {ai_message} "
+                f"is not a valid json format, json error: {error}. "
+                f"You need to strictly return the raw JSON format. ",
+            )
+
         try:
             param: ToolInput = self._input_convert(ai_message, ToolInput)
         except Exception as e:
@@ -100,6 +111,16 @@ async def run(
                 content="The requested correctly structured answer could not be found.",
             )
 
+        if param.tool_name is None or param.tool_name == "":
+            # can not choice tools， it must be some reason
+            return ActionOutput(
+                is_exe_success=False,
+                # content= param.thought,
+                content=f"There are no suitable tools available "
+                f"to achieve the user's goal: '{param.thought}'",
+                have_retry=False,
+            )
+
         try:
             tool_packs = ToolPack.from_resource(self.resource)
             if not tool_packs:
@@ -137,10 +158,25 @@ async def run(
                 is_exe_success=response_success,
                 content=str(tool_result),
                 view=view,
+                thoughts=param.thought,
+                action=str({"tool_name": param.tool_name, "args": param.args}),
                 observations=str(tool_result),
             )
         except Exception as e:
             logger.exception("Tool Action Run Failed！")
             return ActionOutput(
-                is_exe_success=False, content=f"Tool action run failed!{str(e)}"
+                is_exe_success=False,
+                content=f"Tool action run failed!{str(e)}",
+                action=str({"tool_name": param.tool_name, "args": param.args}),
             )
+
+
+def parse_json_safe(json_str):
+    """Try to parse json."""
+    try:
+        # try to parse json
+        data = json.loads(json_str)
+        return True, data
+    except json.JSONDecodeError as e:
+        # 捕捉JSON解析错误并返回详细信息
+        return False, e.msg

dbgpt/agent/expand/tool_assistant_agent.py

@@ -1,9 +1,12 @@
 """Plugin Assistant Agent."""
 
 import logging
+from typing import List, Optional
 
+from .. import Resource, ResourceType
 from ..core.base_agent import ConversableAgent
 from ..core.profile import DynConfig, ProfileConfig
+from ..resource import BaseTool
 from .actions.tool_action import ToolAction
 
 logger = logging.getLogger(__name__)
@@ -37,7 +40,10 @@ class ToolAssistantAgent(ConversableAgent):
                 "goal.",
                 "Please output the selected tool name and specific parameter "
                 "information in json format according to the following required format."
-                " If there is an example, please refer to the sample format output.",
+                "If there is an example, please refer to the sample format output.",
+                "It is not necessarily required to select a tool for execution. "
+                "If the tool to be used or its parameters cannot be clearly "
+                "determined based on the user's input, you can choose not to execute.",
             ],
             category="agent",
             key="dbgpt_agent_expand_plugin_assistant_agent_constraints",
@@ -54,3 +60,40 @@ def __init__(self, **kwargs):
         """Create a new instance of ToolAssistantAgent."""
         super().__init__(**kwargs)
         self._init_actions([ToolAction])
+
+    @property
+    def desc(self) -> Optional[str]:
+        """Return desc of this agent."""
+        tools = _get_tools_by_resource(self.resource)
+        if tools is None or len(tools) == 0:
+            return "Has no tools to use"
+
+        tools_desc_list = []
+        for i in range(len(tools)):
+            tool = tools[i]
+            s = f"{i + 1}. tool {tool.name}, can {tool.description}."
+            tools_desc_list.append(s)
+
+        return (
+            "Can use the following tools to complete the task objectives, "
+            "tool information: "
+            f"{' '.join(tools_desc_list)}"
+        )
+
+
+def _get_tools_by_resource(resource: Optional[Resource]) -> Optional[List[BaseTool]]:
+    tools: List[BaseTool] = []
+
+    if resource is None:
+        return tools
+
+    if resource.type() == ResourceType.Tool and isinstance(resource, BaseTool):
+        tools.append(resource)
+    elif resource.type() == ResourceType.Pack:
+        for sub_res in resource.sub_resources:
+            res_list = _get_tools_by_resource(sub_res)
+            if res_list is not None and len(res_list) > 0:
+                for res in res_list:
+                    tools.append(res)
+
+    return tools