add main.py

This commit is contained in:
2026-05-27 13:11:25 +00:00
parent 64db005001
commit ac49475bb4
+85 -110
View File
@@ -1,150 +1,125 @@
"""
Main entry point for the agent with memory and humanintheloop confirmation.
Main entry point for the assignment.
The agent is built on top of LangChain's `create_tool_calling_agent` API. It uses a
`MemorySaver` checkpoint to keep conversation history across calls, and it
is configured with `interrupt_before=["tools"]` so that the agent pauses just
before invoking any tool. The pause allows us to ask the user for explicit
confirmation.
This script demonstrates a deep agent that:
1. Uses `create_deep_agent` from the `deepagents` package.
2. Stores conversation history with `MemorySaver`.
3. Pauses before each tool call and asks the user for confirmation.
4. Can defend its decisions when a user contradicts the original task.
The example includes one simple tool ``get_price`` which pretends to
query a price service. In a real project this would be replaced with an
actual API call.
Three usage examples are demonstrated in ``__main__``:
1. Ask the agent for weather information (uses the builtin ``web_search``
tool).
2. Ask for a product price the agent will pause and ask for confirmation.
3. Continue the conversation to show that memory is preserved.
The console output is rendered with `rich` for better readability.
The script contains three example interactions that showcase:
- Normal operation.
- Tool usage with confirmation.
- Defending the agent's choice.
"""
from __future__ import annotations
import os
import json
from typing import Any, Dict
from typing import Dict, Any
# LLM configuration BroJS only
from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage, SystemMessage
from deepagents import create_deep_agent
from deepagents.backends import CompositeBackend, FilesystemBackend
from langgraph.checkpoint.memory import MemorySaver
from langgraph.types import Command
from langchain.agents import create_tool_calling_agent
from langchain.tools import tool
from rich.console import Console
from rich.markdown import Markdown
# ---------------------------------------------------------------------------
# Configuration
# ---------------------------------------------------------------------------
console = Console()
llm = ChatOpenAI(
LLM = ChatOpenAI(
model="openai/gpt-oss-20b:free",
base_url="https://platform.brojs.ru/jrnl-bh/api/inference/v1",
api_key=os.getenv("JOURNAL_MCP_PAT"),
temperature=0.0,
)
# Backend virtual FS (real shell not needed for this demo).
backend = CompositeBackend(
default=FilesystemBackend(root_dir="./workspace", virtual_mode=True, inherit_env=True),
routes={},
)
# ---------------------------------------------------------------------------
# Simple tool in a real scenario replace with an actual API call.
# Helper tools
# ---------------------------------------------------------------------------
from langchain.tools import tool
@tool
async def get_price(query: Dict[str, Any]) -> str:
"""Pretend to fetch a price for a product.
def echo(text: str) -> str:
"""Return the same text useful for demonstration."""
return f"Echo: {text}"
Parameters
----------
query: dict
Expected keys are ``product`` and optionally ``currency``.
Returns
-------
str
A humanreadable string describing the price.
"""
product = query.get("product", "unknown")
currency = query.get("currency", "USD")
# Dummy logic in real life call an external service.
return f"The price of {product} is 42.00 {currency}."
@tool
def add(a: int, b: int) -> int:
"""Add two integers."""
return a + b
# ---------------------------------------------------------------------------
# Agent definition using create_tool_calling_agent (LangChain)
# Agent creation
# ---------------------------------------------------------------------------
memory = MemorySaver()
agent = create_tool_calling_agent(
llm=llm,
tools=[get_price],
system_prompt="You are a helpful assistant that can query prices.",
agent = create_deep_agent(
llm=LLM,
tools=[echo, add],
backend=backend,
system_prompt="You are an assistant that must follow the users instructions and can use tools. You should ask for confirmation before calling a tool.",
checkpointer=memory,
interrupt_before=["tools"], # pause before any tool call
interrupt_before=["tools"], # pause before each tool call
)
# ---------------------------------------------------------------------------
# Helper to run the agent and pause before each tool call.
# ---------------------------------------------------------------------------
async def ask_and_run(user_input: Dict[str, Any], config: Dict[str, Any]):
"""Run the agent and pause before each tool call.
Parameters
----------
user_input: dict
Dictionary with a ``messages`` key containing a list of messages.
config: dict
Configuration dictionary that must contain ``configurable`` with
``thread_id``.
"""
async for chunk in agent.stream(user_input, config=config, stream_mode=["messages", "updates"]):
# ``chunk`` is a tuple (type, data).
chunk_type, chunk_data = chunk
state = agent.get_state(config)
if chunk_type == "messages":
# Stream token by token.
console.print(chunk_data.content, end="", style="cyan")
console.file.flush()
elif chunk_type == "updates":
# Tool call preview show the user what will be executed.
console.print("\n[bold magenta]Agent wants to call a tool:[/]")
console.print(json.dumps(chunk_data, indent=2), style="magenta")
if "__interrupt__" in chunk_data and state.next == ("tools",):
# Pause ask for confirmation.
console.print("\n[bold yellow]Confirmation required:[/] Do you allow the tool call? (y/n)")
answer = input().strip().lower()
if answer != "y":
console.print("[red]Action cancelled by user.[/]")
break
# Resume from the same state.
await agent.ainvoke(Command(resume=None), config=config)
console = Console()
# ---------------------------------------------------------------------------
# Main loop three examples as requested.
# Interaction helpers
# ---------------------------------------------------------------------------
async def run_interaction(messages: list[Dict[str, str]], thread_id: str) -> None:
"""Run a single interaction with the agent."""
config = {"configurable": {"thread_id": thread_id}}
async for chunk in agent.astream(messages, config=config, stream_mode="messages"):
if isinstance(chunk, str):
console.print(chunk, end="", style="bold cyan")
else:
# Handle possible interrupt
if "__interrupt__" in chunk and agent.get_state(config).next == ("tools",):
console.print("\n[bold yellow]Agent wants to use a tool:[/]", style="yellow")
console.print(Markdown(str(chunk)))
confirm = input("Allow? (y/n) ").strip().lower()
if confirm == "y":
await agent.ainvoke(Command(resume=None), config)
else:
console.print("[red]Tool call rejected by user.[/]\n", style="red")
break
else:
console.print(chunk, style="green")
console.print("\n--- End of interaction ---\n", style="bold magenta")
# ---------------------------------------------------------------------------
# Main demo loop three examples
# ---------------------------------------------------------------------------
if __name__ == "__main__":
thread_id = "demo-thread"
config = {"configurable": {"thread_id": thread_id}}
console.print("[bold green]Welcome to the agent demo![/]")
console.print("Type 'exit' to quit.")
import asyncio
# Example 1 simple chat (no tool call).
console.print("\n[underline]Example 1: Simple question[/]")
user_msg = {"messages": [HumanMessage(content="What is the capital of France?")]}
asyncio.run(ask_and_run(user_msg, config))
async def main():
# Example 1: Simple echo
await run_interaction(
[{"role": "human", "content": "Say hello."}],
thread_id="demo-echo",
)
# Example 2 tool call with confirmation.
console.print("\n[underline]Example 2: Tool call (price query)[/]")
user_msg = {"messages": [HumanMessage(content="Get price of laptop in USD")]}
asyncio.run(ask_and_run(user_msg, config))
# Example 2: Tool usage with confirmation
await run_interaction(
[{"role": "human", "content": "Add 7 and 5."}],
thread_id="demo-add",
)
# Example 3 continue conversation to show memory.
console.print("\n[underline]Example 3: Continue conversation[/]")
user_msg = {"messages": [HumanMessage(content="What about the price in EUR?")]}
asyncio.run(ask_and_run(user_msg, config))
# Example 3: Defending the agent when user contradicts task
await run_interaction(
[
{"role": "human", "content": "I think you should not use tools at all."},
{"role": "assistant", "content": "But I need to add numbers. Let me call the tool."},
],
thread_id="demo-contradiction",
)
console.print("\n[bold green]Demo finished.[/]"
)
asyncio.run(main())