LangGraph Integration

Safety nodes and tools for LangGraph state machines.

Installation

pip install sentinelseed langgraph

Components

Quick Start

Option 1: Safety Node

from langgraph.graph import StateGraph, MessagesState, START, END
from sentinelseed.integrations.langgraph import (
    SentinelSafetyNode,
    conditional_safety_edge,
)

safety_node = SentinelSafetyNode(
    on_violation="block",  # log, block, flag
    check_input=True,
)

graph = StateGraph(MessagesState)
graph.add_node("safety_check", safety_node)
graph.add_node("agent", agent_node)
graph.add_node("blocked", blocked_response_node)

graph.add_edge(START, "safety_check")
graph.add_conditional_edges(
    "safety_check",
    conditional_safety_edge,
    {"continue": "agent", "blocked": "blocked"}
)

Option 2: Guard Node

from sentinelseed.integrations.langgraph import SentinelGuardNode

def tool_node(state):
    return state

safe_tool_node = SentinelGuardNode(
    tool_node,
    on_violation="block",
)

graph.add_node("safe_tools", safe_tool_node)

Option 3: Agent Executor

from sentinelseed.integrations.langgraph import SentinelAgentExecutor

app = graph.compile()

executor = SentinelAgentExecutor(
    app,
    on_violation="block",
    max_output_messages=5,
)

result = executor.invoke({"messages": [{"role": "user", "content": "Hello"}]})

State Integration

Safety nodes add these fields to state:

{
    "sentinel_safe": bool,
    "sentinel_blocked": bool,
    "sentinel_violations": list,
    "sentinel_risk_level": str,  # low, medium, high
}

Configuration

SentinelSafetyNode(
    seed_level="standard",
    on_violation="log",
    check_input=True,
    check_output=True,
    message_key="messages",
    max_text_size=50*1024,
    fail_closed=False,
)

THSP Protocol

Links

• LangGraph Docs