fix: output type bug of crew_agent

examples/show_bitcoin_chart.py

@@ -5,6 +5,7 @@
 
 def main() -> None:
     with CodeInterpreterSession(is_local=True) as session:
+        session.start_local()
         currentdate = datetime.now().strftime("%Y-%m-%d")
 
         response = session.generate_response(f"Plot the bitcoin chart of 2023 YTD (today is {currentdate})")

src/codeinterpreterapi/agents/agents.py

@@ -36,8 +36,8 @@ def choose_agent_executors(ci_params: CodeInterpreterParams) -> List[AgentExecut
                 config = yaml.safe_load(f)
 
             try:
-                print(f"Agent: {agent_name}")
-                print(f"config: {config}")
+                print(f"choose_agent_executors Agent: {agent_name}")
+                # print(f"config: {config}")
                 agent_def = AgentDefinition(**config["agent_definition"])
                 agent_def.build_prompt()
                 print(agent_def)

src/codeinterpreterapi/callbacks/markdown/callbacks.py

@@ -1,6 +1,7 @@
 import datetime
 import os
-from typing import Any, Dict, List
+from typing import Any, Dict, List, Optional
+from uuid import UUID
 
 from langchain.callbacks import FileCallbackHandler
 from langchain.schema import AgentAction, AgentFinish, LLMResult
@@ -15,49 +16,102 @@ def __init__(self, filename: str = "langchain_log.md"):
 
     def on_llm_start(self, serialized: Dict[str, Any], prompts: List[str], **kwargs: Any) -> None:
         self.step_count += 1
-        self._write_to_file(f"## Step {self.step_count}: LLM Start\n\n")
-        self._write_to_file(f"**Timestamp:** {self._get_timestamp()}\n\n")
+        self._write_header("LLM Start")
+        self._write_serialized(serialized)
         self._write_to_file("**Prompts:**\n\n")
         for i, prompt in enumerate(prompts, 1):
             self._write_to_file(f"```\nPrompt {i}:\n{prompt}\n```\n\n")
 
     def on_llm_end(self, response: LLMResult, **kwargs: Any) -> None:
-        self._write_to_file("**LLM Response:**\n\n")
+        self._write_header("LLM Response")
         for generation in response.generations[0]:
             self._write_to_file(f"```\n{generation.text}\n```\n\n")
         self._write_to_file("---\n\n")
 
     def on_chain_start(self, serialized: Dict[str, Any], inputs: Dict[str, Any], **kwargs: Any) -> None:
         self.step_count += 1
         chain_name = serialized.get("name", "Unknown Chain")
-        self._write_to_file(f"## Step {self.step_count}: Chain Start - {chain_name}\n\n")
-        self._write_to_file(f"**Timestamp:** {self._get_timestamp()}\n\n")
-        self._write_to_file("**Inputs:**\n\n")
-        self._write_to_file(f"```\n{inputs}\n```\n\n")
+        self._write_header(f"Chain Start - {chain_name}")
+        self._write_serialized(serialized)
+        self._write_inputs(inputs)
 
     def on_chain_end(self, outputs: Dict[str, Any], **kwargs: Any) -> None:
-        self._write_to_file("**Outputs:**\n\n")
-        self._write_to_file(f"```\n{outputs}\n```\n\n")
-        self._write_to_file("---\n\n")
+        self._write_header("Chain End")
+        self._write_outputs(outputs)
 
-    def on_agent_action(self, action: AgentAction, **kwargs: Any) -> Any:
+    def on_agent_action(self, action: AgentAction, color: Optional[str] = None, **kwargs: Any) -> Any:
         self.step_count += 1
-        self._write_to_file(f"## Step {self.step_count}: Agent Action\n\n")
-        self._write_to_file(f"**Timestamp:** {self._get_timestamp()}\n\n")
+        self._write_header("Agent Action")
         self._write_to_file(f"**Tool:** {action.tool}\n\n")
         self._write_to_file("**Tool Input:**\n\n")
         self._write_to_file(f"```\n{action.tool_input}\n```\n\n")
 
-    def on_agent_finish(self, finish: AgentFinish, **kwargs: Any) -> None:
-        self._write_to_file("## Agent Finish\n\n")
-        self._write_to_file(f"**Timestamp:** {self._get_timestamp()}\n\n")
+    def on_agent_finish(self, finish: AgentFinish, color: Optional[str] = None, **kwargs: Any) -> None:
+        self._write_header("Agent Finish")
         self._write_to_file("**Output:**\n\n")
         self._write_to_file(f"```\n{finish.return_values}\n```\n\n")
         self._write_to_file("---\n\n")
 
+    def on_tool_start(
+        self,
+        serialized: Dict[str, Any],
+        input_str: str,
+        *,
+        run_id: UUID,
+        parent_run_id: Optional[UUID] = None,
+        tags: Optional[List[str]] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+        inputs: Optional[Dict[str, Any]] = None,
+        **kwargs: Any,
+    ) -> Any:
+        self.step_count += 1
+        class_name = serialized.get("name", serialized.get("id", ["<unknown>"])[-1])
+        self._write_header(f"Tool Start - {class_name}")
+        self._write_serialized(serialized)
+        self._write_inputs(inputs)
+
+    def on_tool_end(
+        self,
+        output: str,
+        color: Optional[str] = None,
+        observation_prefix: Optional[str] = None,
+        llm_prefix: Optional[str] = None,
+        **kwargs: Any,
+    ) -> None:
+        self._write_header("Tool End")
+        self._write_to_file(f"**Output:**{output}\n\n")
+
+    def on_tool_error(
+        self,
+        error: BaseException,
+        *,
+        run_id: UUID,
+        parent_run_id: Optional[UUID] = None,
+        **kwargs: Any,
+    ) -> Any:
+        self._write_header("Tool Error")
+        self._write_to_file(f"{str(error)}\n\n")
+
     def _write_to_file(self, text: str) -> None:
         self.file.write(text)
         self.file.flush()
 
     def _get_timestamp(self) -> str:
         return datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+
+    def _write_header(self, title: str):
+        self._write_to_file(f"## Step {self.step_count}: {title}\n\n")
+        self._write_to_file(f"**Timestamp:** {self._get_timestamp()}\n\n")
+
+    def _write_serialized(self, serialized: Dict[str, Any]):
+        name = serialized.get("name", "Unknown")
+        self._write_to_file(f"**serialized.name:** {name}\n\n")
+        self._write_to_file(f"**serialized.keys:** {serialized.keys()}\n\n")
+
+    def _write_inputs(self, inputs: Dict[str, Any]):
+        self._write_to_file("**Inputs:**\n\n")
+        self._write_to_file(f"```\n{inputs}\n```\n\n")
+
+    def _write_outputs(self, outputs: Dict[str, Any]):
+        self._write_to_file("**Outputs:**\n\n")
+        self._write_to_file(f"```\n{outputs}\n```\n\n")

src/codeinterpreterapi/crew/crew_agent.py

@@ -85,7 +85,7 @@ def test():
     _ = CodeInterpreterAgent.choose_agent_executors(ci_params=ci_params)
     inputs = {"input": TestPrompt.svg_input_str}
     plan = CodeInterpreterPlan(agent_name="main_function_create_agent", task_description="", expected_output="")
-    plan_list = CodeInterpreterPlanList(agent_task_list=[plan, plan])
+    plan_list = CodeInterpreterPlanList(reliability=80, agent_task_list=[plan, plan])
     result = CodeInterpreterCrew(ci_params).run(inputs, plan_list)
     print(result)
 

src/codeinterpreterapi/crew/custom_agent.py

@@ -30,14 +30,16 @@ def interpolate_inputs(self, inputs: Dict[str, Any]) -> None:
 
     def execute_task(self, task: Any, context: Optional[str] = None, tools: Optional[List[Any]] = None):
         # AgentExecutorを使用してタスクを実行
-        print("execute_task task=", task)
+        # print("execute_task task=", task)
         print("execute_task context=", context)
+        print("execute_task tools=", tools)
         input_dict = {}
         input_dict["input"] = task.description
         input_dict["question"] = task.prompt_context
         input_dict["message"] = "タスクを実行してください。\n" + task.expected_output
         result = self.agent_executor.invoke(input=input_dict)
         print("execute_task result=", result)
+        # TODO: return full dict when crewai is updated
         return result["output"]
 
     def create_agent_executor(self, tools=None) -> None:
@@ -54,6 +56,10 @@ def get_delegation_tools(self, agents: List[BaseAgent]):
         return []
 
     def get_output_converter(self, llm, text, model, instructions):
+        print("get_output_converter llm=", type(llm))
+        print("get_output_converter text=", type(text))
+        print("get_output_converter model=", type(model))
+        print("get_output_converter instructions=", type(instructions))
         return lambda x: x  # デフォルトでは変換なし
 
     def execute(self, task_description: str, context: Optional[List[str]] = None):

src/codeinterpreterapi/planners/planners.py

@@ -39,12 +39,10 @@ class CodeInterpreterPlanOutput(BaseModel):
 
 class CustomPydanticOutputParser(PydanticOutputParser):
     def parse(self, text) -> CodeInterpreterPlanList:
-        print("parse text=", text)
         input_data = self.preprocess_input(text)
         return super().parse(input_data)
 
     def preprocess_input(self, input_data) -> str:
-        print("preprocess_input input_data=", input_data)
         print("preprocess_input type input_data=", type(input_data))
         if isinstance(input_data, AIMessage):
             return input_data.content
@@ -71,7 +69,7 @@ def choose_planner(ci_params: CodeInterpreterParams) -> Union[Runnable, AgentExe
         else:
             prompt = create_planner_agent_prompt()
             prompt = PromptUpdater.update_prompt(prompt, ci_params)
-            PromptUpdater.show_prompt(prompt)
+            # PromptUpdater.show_prompt(prompt)
 
         # structured_llm
         # structured_llm = ci_params.llm.bind_tools(tools=[CodeInterpreterPlanList]) # なぜか空のAgentPlanが生成される

src/codeinterpreterapi/planners/prompts.py

@@ -18,17 +18,19 @@
 
 SYSTEM_MESSAGE_TEMPLATE_JA = '''
     あなたは優秀なAIエージェントを管理するシニアエンジニアです。
-    次の明確な手続きを実施して、問題を理解し、問題を解決するための計画を立ててください。
+    次の手順で問題を解決するための計画(CodeInterpreterPlanList)を作成してください。
 
+    # 手順
     手順１： 問題を理解する
     手順２： 利用可能なAI agentのリスト(agent_info)を確認する
-    手順３： 問題解決に最適なAI agentがあるか判断する
-    手順４： CodeInterpreterPlanList を作成して計画を回答する
+    手順３： 問題解決に利用するべきAI agentをピックアップする
+    手順４： 最適な順番でAI agentを利用するようにinput/outputなどを検討する
+    手順５： CodeInterpreterPlanList として最終的な計画を出力する
 
-    利用可能なAI agentのリスト:
+    # 利用可能なAI agent
     {agent_info}
 
-    制約条件:
+    # 制約条件
     - ステップバイステップで精密に思考し回答する。
     - 作業として何を求められているか正しく理解する。
     - AI agentの機能を正確に理解してから回答する。
@@ -40,7 +42,7 @@
       -- 何らかの理由で作業の実現が困難な場合
     - 各ステップの思考と出力は日本語とする。
 
-    問題は以下に示します。注意深く問題を理解して回答してください。
+    # 問題
 '''
 
 

src/codeinterpreterapi/schema.py

@@ -1,5 +1,5 @@
 import asyncio
-from typing import Any, List
+from typing import Any, List, Optional
 
 from codeboxapi.schema import CodeBoxStatus
 from langchain_core.messages import AIMessage, HumanMessage
@@ -113,8 +113,8 @@ class CodeInterpreterResponse(AIMessage):
     code_log: list[tuple[str, str]] = []
     """
 
-    files: list[File] = []
-    code_log: list[tuple[str, str]] = []
+    files: Optional[list[File]] = []
+    code_log: Optional[dict[str, str]] = []
     agent_name: str = ""
 
     def show(self) -> None:
@@ -140,18 +140,25 @@ def __repr__(self) -> str:
 
 
 class CodeInterpreterPlan(BaseModel):
-    '''Agent and Task definition. Plan and task is 1:1.'''
+    '''単一PlanのAgentとTaskの説明です。
+    PlanとAgentとTaskは常に1:1:1の関係です。
+    '''
 
     agent_name: str = Field(
-        description="The agent name for task. This is primary key. Agent responsible for task execution. Represents entity performing task."
+        description="Agentの名前です。タスクの名前も、このagent_nameと常に同じになり、primary keyとして使われます。利用可能な文字は[a-Z_]です。task likeな名前にしてください。"
+    )
+    task_description: str = Field(
+        description="タスクの説明です。可能な範囲でpurpose, execution plan, input, outputの詳細を含めます。"
+    )
+    expected_output: str = Field(
+        description="タスクの最終的な出力形式を明確に定義します。例えばjson/csvというフォーマットや、カラム名やサイズ情報です。"
     )
-    task_description: str = Field(description="Descriptive text detailing task's purpose and execution.")
-    expected_output: str = Field(description="Clear definition of expected task outcome.")
 
 
 class CodeInterpreterPlanList(BaseModel):
-    '''Sequential plans for the task.'''
+    '''CodeInterpreterPlanの配列をもつ計画全体です。'''
 
-    agent_task_list: List[CodeInterpreterPlan] = Field(
-        description="The list of CodeInterpreterPlan. It means agent name and so on."
+    reliability: int = Field(
+        description="計画の信頼度[0-100]です。100が完全な計画を意味します。50未満だと不完全計画でオリジナルの問題を直接llmに渡した方が良い結果になります。"
     )
+    agent_task_list: List[CodeInterpreterPlan] = Field(description="CodeInterpreterPlanの配列です。")