datagusto-sdk 0.1.7

Datagusto safety middleware for LangChain/LangGraph agents

Datagusto SDK (Safety Middleware)

DatagustoSafetyMiddleware is a LangChain/LangGraph middleware that:

• Extracts tool definitions from requests and (optionally) registers them to your Datagusto backend.
• Submits user instructions for alignment.
• Calls safety validation before and after tool execution, blocking execution when the backend returns should_proceed = false.

This package is published as datagusto-sdk and exposes DatagustoSafetyMiddleware.

Features

• Tool definition extraction (names, descriptions, JSON schemas for input/output).
• Automatic registration diffing (skips duplicate payloads).
• Alignment submission for the latest user instruction.
• Tool-level validation on start/end with structured payloads.
• Verbose logging for request/response bodies (optional).

Requirements

• Python >= 3.12
• Dependencies: langchain, langgraph

Installation

pip install datagusto-sdk

Quickstart (LangChain agent)

from langchain.agents import create_agent
from langchain_core.tools import tool
from langchain.messages import HumanMessage
from datagusto_sdk import DatagustoSafetyMiddleware

@tool
def ping(msg: str) -> str:
    return f"pong: {msg}"

middleware = DatagustoSafetyMiddleware(
    verbose=True,
    include_schema=True,
    tool_filter=None,  # optional: list of tool names to register
)

agent = create_agent(
    model="claude-3-5-sonnet-latest",
    tools=[ping],
    middleware=[middleware],
)

agent.invoke({"messages": [HumanMessage("ping please")]})

Configuration

• Environment variables (read automatically if args not provided):
  • SERVER_URL: Base URL for Datagusto backend (e.g., https://api.example.com)
  • API_KEY: Bearer token for API calls
• Runtime options:
  • verbose (bool): log request/response bodies; default True.
  • include_schema (bool): include JSON schemas for tool IO; default True.
  • tool_filter (list[str] | None): only register tools with names in this list.

Behavior overview

1. wrap_model_call
   • Extract tools from request.tools, build payload, hash it.
   • If server_url and api_key are set and payload hash changed, call register API.
   • Extract latest human message and (if changed) post alignment; store session_id for validation.
2. wrap_tool_call
   • On each tool call, if session_id exists, call validate API with on_start.
   • If backend responds should_proceed = false, raises RuntimeError to block.
   • After tool runs, call validate API with on_end; may block further processing similarly.

Error handling

• Network / HTTP errors are logged (when verbose=True) but do not crash the agent unless the backend explicitly blocks via should_proceed = false.
• Validation blocks raise RuntimeError with the backend response for transparency.

License

MIT License. See LICENSE.