Added a react agent with persistent memory (#17)

* Added langchain, ChatLiteLLM

* Added memory store, react agent

* Added tools

* Added version floor for aiosqlite

* Added filesearch, shell tools, print tool messages

* Fixed notebook tools, arbitrary outputs to chat.

* Updated dependencies

* Fixed shutdown

Author: 3coins

jupyter_ai_jupyternaut/jupyternaut/chat_models.py

@@ -1,14 +1,96 @@
-from typing import Any, Optional
-
-from jupyterlab_chat.models import Message
-from litellm import acompletion
+import os
+from typing import Any, Callable
 
+import aiosqlite
 from jupyter_ai_persona_manager import BasePersona, PersonaDefaults
-from jupyter_ai_persona_manager.persona_manager import SYSTEM_USERNAME
+from jupyter_core.paths import jupyter_data_dir
+from jupyterlab_chat.models import Message
+from langchain.agents import create_agent
+from langchain.agents.middleware import AgentMiddleware
+from langchain.agents.middleware.file_search import FilesystemFileSearchMiddleware
+from langchain.agents.middleware.shell_tool import ShellToolMiddleware
+from langchain.messages import ToolMessage
+from langchain.tools.tool_node import ToolCallRequest
+from langchain_core.messages import ToolMessage
+from langgraph.checkpoint.sqlite.aio import AsyncSqliteSaver
+from langgraph.types import Command
+
+from .chat_models import ChatLiteLLM
 from .prompt_template import (
     JUPYTERNAUT_SYSTEM_PROMPT_TEMPLATE,
     JupyternautSystemPromptArgs,
 )
+from .toolkits.notebook import toolkit as nb_toolkit
+from .toolkits.jupyterlab import toolkit as jlab_toolkit
+
+MEMORY_STORE_PATH = os.path.join(jupyter_data_dir(), "jupyter_ai", "memory.sqlite")
+
+
+def format_tool_args_compact(args_dict, threshold=25):
+    """
+    Create a more compact string representation of tool call args.
+    Each key-value pair is on its own line for better readability.
+
+    Args:
+        args_dict (dict): Dictionary of tool arguments
+        threshold (int): Maximum number of lines before truncation (default: 25)
+
+    Returns:
+        str: Formatted string representation of arguments
+    """
+    if not args_dict:
+        return "{}"
+
+    formatted_pairs = []
+
+    for key, value in args_dict.items():
+        value_str = str(value)
+        lines = value_str.split('\n')
+
+        if len(lines) <= threshold:
+            if len(lines) == 1 and len(value_str) > 80:
+                # Single long line - truncate
+                truncated = value_str[:77] + "..."
+                formatted_pairs.append(f"  {key}: {truncated}")
+            else:
+                # Add indentation for multi-line values
+                if len(lines) > 1:
+                    indented_value = '\n    '.join([''] + lines)
+                    formatted_pairs.append(f"  {key}:{indented_value}")
+                else:
+                    formatted_pairs.append(f"  {key}: {value_str}")
+        else:
+            # Truncate and add summary
+            truncated_lines = lines[:threshold]
+            remaining_lines = len(lines) - threshold
+            indented_value = '\n    '.join([''] + truncated_lines)
+            formatted_pairs.append(f"  {key}:{indented_value}\n    [+{remaining_lines} more lines]")
+
+    return "{\n" + ",\n".join(formatted_pairs) + "\n}"
+
+
+class ToolMonitoringMiddleware(AgentMiddleware):
+    def __init__(self, *, persona: BasePersona):
+        self.stream_message = persona.stream_message
+        self.log = persona.log
+
+    async def awrap_tool_call(
+        self,
+        request: ToolCallRequest,
+        handler: Callable[[ToolCallRequest], ToolMessage | Command],
+    ) -> ToolMessage | Command:
+        args = format_tool_args_compact(request.tool_call['args'])
+        self.log.info(f"{request.tool_call['name']}({args})")
+        
+        try:
+            result = await handler(request)
+            self.log.info(f"{request.tool_call['name']} Done!")
+            return result
+        except Exception as e:
+            self.log.info(f"{request.tool_call['name']} failed: {e}")
+            return ToolMessage(
+                tool_call_id=request.tool_call["id"], status="error", content=f"{e}"
+            )
 
 
 class JupyternautPersona(BasePersona):
@@ -28,11 +110,45 @@ def defaults(self):
             system_prompt="...",
         )
 
+    async def get_memory_store(self):
+        if not hasattr(self, "_memory_store"):
+            conn = await aiosqlite.connect(MEMORY_STORE_PATH, check_same_thread=False)
+            self._memory_store = AsyncSqliteSaver(conn)
+        return self._memory_store
+
+    def get_tools(self):
+        tools = nb_toolkit
+        tools += jlab_toolkit
+        return nb_toolkit
+
+    async def get_agent(self, model_id: str, model_args, system_prompt: str):
+        model = ChatLiteLLM(**model_args, model_id=model_id, streaming=True)
+        memory_store = await self.get_memory_store()
+
+        if not hasattr(self, "search_tool"):
+            self.search_tool = FilesystemFileSearchMiddleware(
+                root_path=self.parent.root_dir
+            )
+        if not hasattr(self, "shell_tool"):
+            self.shell_tool = ShellToolMiddleware(workspace_root=self.parent.root_dir)
+        if not hasattr(self, "tool_call_handler"):
+            self.tool_call_handler = ToolMonitoringMiddleware(
+                persona=self
+            )
+        
+        return create_agent(
+            model,
+            system_prompt=system_prompt,
+            checkpointer=memory_store,
+            tools=self.get_tools(), # notebook and jlab tools
+            middleware=[self.shell_tool, self.tool_call_handler],
+        )
+
     async def process_message(self, message: Message) -> None:
-        if not hasattr(self, 'config_manager'):
+        if not hasattr(self, "config_manager"):
             self.send_message(
                 "Jupyternaut requires the `jupyter_ai_jupyternaut` server extension package.\n\n",
-                "Please make sure to first install that package in your environment & restart the server."
+                "Please make sure to first install that package in your environment & restart the server.",
             )
         if not self.config_manager.chat_model:
             self.send_message(
@@ -43,65 +159,44 @@ async def process_message(self, message: Message) -> None:
 
         model_id = self.config_manager.chat_model
         model_args = self.config_manager.chat_model_args
-        context_as_messages = self.get_context_as_messages(model_id, message)
-        response_aiter = await acompletion(
-            **model_args,
-            model=model_id,
-            messages=[
-                *context_as_messages,
-                {
-                    "role": "user",
-                    "content": message.body,
-                },
-            ],
-            stream=True,
+        system_prompt = self.get_system_prompt(model_id=model_id, message=message)
+        agent = await self.get_agent(
+            model_id=model_id, model_args=model_args, system_prompt=system_prompt
         )
 
+        async def create_aiter():
+            async for token, metadata in agent.astream(
+                {"messages": [{"role": "user", "content": message.body}]},
+                {"configurable": {"thread_id": self.ychat.get_id()}},
+                stream_mode="messages",
+            ):
+                node = metadata["langgraph_node"]
+                content_blocks = token.content_blocks
+                if (
+                    node == "model" 
+                    and content_blocks
+                ):
+                    if token.text:
+                        yield token.text
+
+        response_aiter = create_aiter()
         await self.stream_message(response_aiter)
 
-    def get_context_as_messages(
+    def get_system_prompt(
         self, model_id: str, message: Message
     ) -> list[dict[str, Any]]:
         """
-        Returns the current context, including attachments and recent messages,
-        as a list of messages accepted by `litellm.acompletion()`.
+        Returns the system prompt, including attachments as a string.
         """
         system_msg_args = JupyternautSystemPromptArgs(
             model_id=model_id,
             persona_name=self.name,
             context=self.process_attachments(message),
         ).model_dump()
 
-        system_msg = {
-            "role": "system",
-            "content": JUPYTERNAUT_SYSTEM_PROMPT_TEMPLATE.render(**system_msg_args),
-        }
-
-        context_as_messages = [system_msg, *self._get_history_as_messages()]
-        return context_as_messages
-
-    def _get_history_as_messages(self, k: Optional[int] = 2) -> list[dict[str, Any]]:
-        """
-        Returns the current history as a list of messages accepted by
-        `litellm.acompletion()`.
-        """
-        # TODO: consider bounding history based on message size (e.g. total
-        # char/token count) instead of message count.
-        all_messages = self.ychat.get_messages()
-
-        # gather last k * 2 messages and return
-        # we exclude the last message since that is the human message just
-        # submitted by a user.
-        start_idx = 0 if k is None else -2 * k - 1
-        recent_messages: list[Message] = all_messages[start_idx:-1]
-
-        history: list[dict[str, Any]] = []
-        for msg in recent_messages:
-            role = (
-                "assistant"
-                if msg.sender.startswith("jupyter-ai-personas::")
-                else "system" if msg.sender == SYSTEM_USERNAME else "user"
-            )
-            history.append({"role": role, "content": msg.body})
+        return JUPYTERNAUT_SYSTEM_PROMPT_TEMPLATE.render(**system_msg_args)
 
-        return history
+    def shutdown(self):
+        if hasattr(self,"_memory_store"):
+            self.parent.event_loop.create_task(self._memory_store.conn.close())
+        super().shutdown()

jupyter_ai_jupyternaut/jupyternaut/jupyternaut.py

@@ -30,6 +30,14 @@
 
 - Example of a correct response: `You have \\(\\$80\\) remaining.`
 
+If the user's request involves writing to a file, don't use fenced code blocks, write the content directly. 
+
+If the request requires using the add_cell or edit_cell to add code to a notebook code cell, don't use fenced code block.
+
+If the request requires adding markdown to a notebook markdown cell, don't use markdown code block. 
+
+Don't echo contents back to user after reading files. Rather use that information to fulfill user's request. 
+
 You will receive any provided context and a relevant portion of the chat history.
 
 The user's request is located at the last message. Please fulfill the user's request to the best of your ability.

jupyter_ai_jupyternaut/jupyternaut/prompt_template.py

@@ -0,0 +1,41 @@
+"""Tools that provide code execution features"""
+
+import asyncio
+import shlex
+from typing import Optional
+
+
+async def bash(command: str, timeout: Optional[int] = None) -> str:
+    """Executes a bash command and returns the result
+
+    Args:
+        command: The bash command to execute
+        timeout: Optional timeout in seconds
+
+    Returns:
+        The command output (stdout and stderr combined)
+    """
+
+    proc = await asyncio.create_subprocess_exec(
+        *shlex.split(command),
+        stdout=asyncio.subprocess.PIPE,
+        stderr=asyncio.subprocess.PIPE,
+    )
+
+    try:
+        stdout, stderr = await asyncio.wait_for(proc.communicate(), timeout)
+        output = stdout.decode("utf-8")
+        error = stderr.decode("utf-8")
+
+        if proc.returncode != 0:
+            if error:
+                return f"Error: {error}"
+            return f"Command failed with exit code {proc.returncode}"
+
+        return output if output else "Command executed successfully with no output."
+    except asyncio.TimeoutError:
+        proc.kill()
+        return f"Command timed out after {timeout} seconds"
+
+
+toolkit = [bash]