Feature/use converter instead of manually trimming (#894)

* Exploring output being passed to tool selector to see if we can better format data

* WIP. Adding JSON repair functionality

* Almost done implementing JSON repair. Testing fixes vs current base case.

* More action cleanup with additional tests

* WIP. Trying to figure out what is going on with tool descriptions

* Update tool description generation

* WIP. Trying to find out what is causing the tools to duplicate

* Replacing tools properly instead of duplicating them accidentally

* Fixing issues for MR

* Update dependencies for JSON_REPAIR

* More cleaning up pull request

* preppering for call

* Fix type-checking issues

---------

Co-authored-by: João Moura <joaomdmoura@gmail.com>

src/crewai/agent.py

@@ -1,9 +1,10 @@
 import os
+from inspect import signature
 from typing import Any, List, Optional, Tuple
 
 from langchain.agents.agent import RunnableAgent
+from langchain.agents.tools import BaseTool
 from langchain.agents.tools import tool as LangChainTool
-from langchain.tools.render import render_text_description
 from langchain_core.agents import AgentAction
 from langchain_core.callbacks import BaseCallbackHandler
 from langchain_openai import ChatOpenAI
@@ -167,14 +168,16 @@ class Agent(BaseAgent):
             if memory.strip() != "":
                 task_prompt += self.i18n.slice("memory").format(memory=memory)
 
-        tools = tools or self.tools
-
-        parsed_tools = self._parse_tools(tools or [])  # type: ignore # Argument 1 to "_parse_tools" of "Agent" has incompatible type "list[Any] | None"; expected "list[Any]"
+        tools = tools or self.tools or []
+        parsed_tools = self._parse_tools(tools)
         self.create_agent_executor(tools=tools)
         self.agent_executor.tools = parsed_tools
         self.agent_executor.task = task
 
-        self.agent_executor.tools_description = render_text_description(parsed_tools)
+        # TODO: COMPARE WITH ARGS AND WITHOUT ARGS
+        self.agent_executor.tools_description = self._render_text_description_and_args(
+            parsed_tools
+        )
         self.agent_executor.tools_names = self.__tools_names(parsed_tools)
 
         if self.crew and self.crew._train:
@@ -189,6 +192,7 @@ class Agent(BaseAgent):
                 "tools": self.agent_executor.tools_description,
             }
         )["output"]
+
         if self.max_rpm:
             self._rpm_controller.stop_rpm_counter()
 
@@ -220,7 +224,7 @@ class Agent(BaseAgent):
         Returns:
             An instance of the CrewAgentExecutor class.
         """
-        tools = tools or self.tools
+        tools = tools or self.tools or []
 
         agent_args = {
             "input": lambda x: x["input"],
@@ -315,6 +319,7 @@ class Agent(BaseAgent):
             tools_list = []
             for tool in tools:
                 tools_list.append(tool)
+
         return tools_list
 
     def _training_handler(self, task_prompt: str) -> str:
@@ -341,6 +346,52 @@ class Agent(BaseAgent):
                 )
         return task_prompt
 
+    def _render_text_description(self, tools: List[BaseTool]) -> str:
+        """Render the tool name and description in plain text.
+
+        Output will be in the format of:
+
+        .. code-block:: markdown
+
+            search: This tool is used for search
+            calculator: This tool is used for math
+        """
+        description = "\n".join(
+            [
+                f"Tool name: {tool.name}\nTool description:\n{tool.description}"
+                for tool in tools
+            ]
+        )
+
+        return description
+
+    def _render_text_description_and_args(self, tools: List[BaseTool]) -> str:
+        """Render the tool name, description, and args in plain text.
+
+        Output will be in the format of:
+
+        .. code-block:: markdown
+
+            search: This tool is used for search, args: {"query": {"type": "string"}}
+            calculator: This tool is used for math, \
+    args: {"expression": {"type": "string"}}
+        """
+        tool_strings = []
+        for tool in tools:
+            args_schema = str(tool.args)
+            if hasattr(tool, "func") and tool.func:
+                sig = signature(tool.func)
+                description = (
+                    f"Tool Name: {tool.name}{sig}\nTool Description: {tool.description}"
+                )
+            else:
+                description = (
+                    f"Tool Name: {tool.name}\nTool Description: {tool.description}"
+                )
+            tool_strings.append(f"{description}\nTool Arguments: {args_schema}")
+
+        return "\n".join(tool_strings)
+
     @staticmethod
     def __tools_names(tools) -> str:
         return ", ".join([t.name for t in tools])

src/crewai/agents/agent_builder/utilities/base_agent_tool.py

@@ -24,6 +24,7 @@ class BaseAgentTools(BaseModel, ABC):
             is_list = coworker.startswith("[") and coworker.endswith("]")
             if is_list:
                 coworker = coworker[1:-1].split(",")[0]
+
         return coworker
 
     def delegate_work(
@@ -40,11 +41,13 @@ class BaseAgentTools(BaseModel, ABC):
         coworker = self._get_coworker(coworker, **kwargs)
         return self._execute(coworker, question, context)
 
-    def _execute(self, agent: Union[str, None], task: str, context: Union[str, None]):
+    def _execute(
+        self, agent_name: Union[str, None], task: str, context: Union[str, None]
+    ):
         """Execute the command."""
         try:
-            if agent is None:
-                agent = ""
+            if agent_name is None:
+                agent_name = ""
 
             # It is important to remove the quotes from the agent name.
             # The reason we have to do this is because less-powerful LLM's
@@ -53,7 +56,7 @@ class BaseAgentTools(BaseModel, ABC):
             # {"task": "....", "coworker": "....
             # when it should look like this:
             # {"task": "....", "coworker": "...."}
-            agent_name = agent.casefold().replace('"', "").replace("\n", "")
+            agent_name = agent_name.casefold().replace('"', "").replace("\n", "")
 
             agent = [  # type: ignore # Incompatible types in assignment (expression has type "list[BaseAgent]", variable has type "str | None")
                 available_agent
@@ -75,9 +78,9 @@ class BaseAgentTools(BaseModel, ABC):
             )
 
         agent = agent[0]
-        task = Task(  # type: ignore # Incompatible types in assignment (expression has type "Task", variable has type "str")
+        task_with_assigned_agent = Task(  # type: ignore # Incompatible types in assignment (expression has type "Task", variable has type "str")
             description=task,
             agent=agent,
             expected_output="Your best answer to your coworker asking you this, accounting for the context shared.",
         )
-        return agent.execute_task(task, context)  # type: ignore # "str" has no attribute "execute_task"
+        return agent.execute_task(task_with_assigned_agent, context)

src/crewai/agents/agent_builder/utilities/base_output_converter.py

@@ -1,7 +1,7 @@
 from abc import ABC, abstractmethod
 from typing import Any, Optional
 
-from pydantic import BaseModel, Field, PrivateAttr
+from pydantic import BaseModel, Field
 
 
 class OutputConverter(BaseModel, ABC):
@@ -21,7 +21,6 @@ class OutputConverter(BaseModel, ABC):
         max_attempts (int): Maximum number of conversion attempts (default: 3).
     """
 
-    _is_gpt: bool = PrivateAttr(default=True)
     text: str = Field(description="Text to be converted.")
     llm: Any = Field(description="The language model to be used to convert the text.")
     model: Any = Field(description="The model to be used to convert the text.")
@@ -41,7 +40,8 @@ class OutputConverter(BaseModel, ABC):
         """Convert text to json."""
         pass
 
-    @abstractmethod  # type: ignore #  Name "_is_gpt" already defined on line 25
-    def _is_gpt(self, llm):  # type: ignore #  Name "_is_gpt" already defined on line 25
+    @property
+    @abstractmethod
+    def is_gpt(self) -> bool:
         """Return if llm provided is of gpt from openai."""
         pass

src/crewai/agents/executor.py

@@ -1,14 +1,6 @@
 import threading
 import time
-from typing import (
-    Any,
-    Dict,
-    Iterator,
-    List,
-    Optional,
-    Tuple,
-    Union,
-)
+from typing import Any, Dict, Iterator, List, Optional, Tuple, Union
 
 from langchain.agents import AgentExecutor
 from langchain.agents.agent import ExceptionTool
@@ -19,9 +11,7 @@ from langchain_core.tools import BaseTool
 from langchain_core.utils.input import get_color_mapping
 from pydantic import InstanceOf
 
-from crewai.agents.agent_builder.base_agent_executor_mixin import (
-    CrewAgentExecutorMixin,
-)
+from crewai.agents.agent_builder.base_agent_executor_mixin import CrewAgentExecutorMixin
 from crewai.agents.tools_handler import ToolsHandler
 from crewai.tools.tool_usage import ToolUsage, ToolUsageErrorException
 from crewai.utilities import I18N

src/crewai/agents/parser.py

@@ -1,6 +1,7 @@
 import re
 from typing import Any, Union
 
+from json_repair import repair_json
 from langchain.agents.output_parsers import ReActSingleInputOutputParser
 from langchain_core.agents import AgentAction, AgentFinish
 from langchain_core.exceptions import OutputParserException
@@ -48,11 +49,15 @@ class CrewAgentParser(ReActSingleInputOutputParser):
                 raise OutputParserException(
                     f"{FINAL_ANSWER_AND_PARSABLE_ACTION_ERROR_MESSAGE}: {text}"
                 )
-            action = action_match.group(1).strip()
-            action_input = action_match.group(2)
-            tool_input = action_input.strip(" ")
-            tool_input = tool_input.strip('"')
-            return AgentAction(action, tool_input, text)
+            action = action_match.group(1)
+            clean_action = self._clean_action(action)
+
+            action_input = action_match.group(2).strip()
+
+            tool_input = action_input.strip(" ").strip('"')
+            safe_tool_input = self._safe_repair_json(tool_input)
+
+            return AgentAction(clean_action, safe_tool_input, text)
 
         elif includes_answer:
             return AgentFinish(
@@ -87,3 +92,30 @@ class CrewAgentParser(ReActSingleInputOutputParser):
                 llm_output=text,
                 send_to_llm=True,
             )
+
+    def _clean_action(self, text: str) -> str:
+        """Clean action string by removing non-essential formatting characters."""
+        return re.sub(r"^\s*\*+\s*|\s*\*+\s*$", "", text).strip()
+
+    def _safe_repair_json(self, tool_input: str) -> str:
+        UNABLE_TO_REPAIR_JSON_RESULTS = ['""', "{}"]
+
+        # Skip repair if the input starts and ends with square brackets
+        # Explanation: The JSON parser has issues handling inputs that are enclosed in square brackets ('[]').
+        # These are typically valid JSON arrays or strings that do not require repair. Attempting to repair such inputs
+        # might lead to unintended alterations, such as wrapping the entire input in additional layers or modifying
+        # the structure in a way that changes its meaning. By skipping the repair for inputs that start and end with
+        # square brackets, we preserve the integrity of these valid JSON structures and avoid unnecessary modifications.
+        if tool_input.startswith("[") and tool_input.endswith("]"):
+            return tool_input
+
+        # Before repair, handle common LLM issues:
+        # 1. Replace """ with " to avoid JSON parser errors
+
+        tool_input = tool_input.replace('"""', '"')
+
+        result = repair_json(tool_input)
+        if result in UNABLE_TO_REPAIR_JSON_RESULTS:
+            return tool_input
+
+        return str(result)

src/crewai/crew.py

@@ -6,15 +6,15 @@ from typing import Any, Dict, List, Optional, Tuple, Union
 
 from langchain_core.callbacks import BaseCallbackHandler
 from pydantic import (
-    UUID4,
-    BaseModel,
-    ConfigDict,
-    Field,
-    InstanceOf,
-    Json,
-    PrivateAttr,
-    field_validator,
-    model_validator,
+  UUID4,
+  BaseModel,
+  ConfigDict,
+  Field,
+  InstanceOf,
+  Json,
+  PrivateAttr,
+  field_validator,
+  model_validator,
 )
 from pydantic_core import PydanticCustomError
 
@@ -503,7 +503,30 @@ class Crew(BaseModel):
                     agent for agent in self.agents if agent != task.agent
                 ]
                 if len(self.agents) > 1 and len(agents_for_delegation) > 0:
-                    task.tools += task.agent.get_delegation_tools(agents_for_delegation)
+                    delegation_tools = task.agent.get_delegation_tools(
+                        agents_for_delegation
+                    )
+
+                    # Add tools if they are not already in task.tools
+                    for new_tool in delegation_tools:
+                        # Find the index of the tool with the same name
+                        existing_tool_index = next(
+                            (
+                                index
+                                for index, tool in enumerate(task.tools or [])
+                                if tool.name == new_tool.name
+                            ),
+                            None,
+                        )
+                        if not task.tools:
+                            task.tools = []
+
+                        if existing_tool_index is not None:
+                            # Replace the existing tool
+                            task.tools[existing_tool_index] = new_tool 
+                        else:
+                            # Add the new tool
+                            task.tools.append(new_tool)
 
             role = task.agent.role if task.agent is not None else "None"
             self._logger.log("debug", f"== Working Agent: {role}", color="bold_purple")

src/crewai/tools/agent_tools.py

@@ -7,7 +7,7 @@ class AgentTools(BaseAgentTools):
     """Default tools around agent delegation"""
 
     def tools(self):
-        coworkers = f"[{', '.join([f'{agent.role}' for agent in self.agents])}]"
+        coworkers = ", ".join([f"{agent.role}" for agent in self.agents])
         tools = [
             StructuredTool.from_function(
                 func=self.delegate_work,

src/crewai/utilities/converter.py

@@ -2,10 +2,8 @@ import json
 
 from langchain.schema import HumanMessage, SystemMessage
 from langchain_openai import ChatOpenAI
-from pydantic import model_validator
-from crewai.agents.agent_builder.utilities.base_output_converter_base import (
-    OutputConverter,
-)
+
+from crewai.agents.agent_builder.utilities.base_output_converter import OutputConverter
 
 
 class ConverterError(Exception):
@@ -19,15 +17,10 @@ class ConverterError(Exception):
 class Converter(OutputConverter):
     """Class that converts text into either pydantic or json."""
 
-    @model_validator(mode="after")
-    def check_llm_provider(self):
-        if not self._is_gpt(self.llm):
-            self._is_gpt = False
-
     def to_pydantic(self, current_attempt=1):
         """Convert text to pydantic."""
         try:
-            if self._is_gpt:
+            if self.is_gpt:
                 return self._create_instructor().to_pydantic()
             else:
                 return self._create_chain().invoke({})
@@ -41,7 +34,7 @@ class Converter(OutputConverter):
     def to_json(self, current_attempt=1):
         """Convert text to json."""
         try:
-            if self._is_gpt:
+            if self.is_gpt:
                 return self._create_instructor().to_json()
             else:
                 return json.dumps(self._create_chain().invoke({}).model_dump())
@@ -75,5 +68,7 @@ class Converter(OutputConverter):
         )
         return new_prompt | self.llm | parser
 
-    def _is_gpt(self, llm) -> bool:  # type: ignore # BUG? Name "_is_gpt" defined on line 20 hides name from outer scope
-        return isinstance(llm, ChatOpenAI) and llm.openai_api_base is None
+    @property
+    def is_gpt(self) -> bool:
+        """Return if llm provided is of gpt from openai."""
+        return isinstance(self.llm, ChatOpenAI) and self.llm.openai_api_base is None