langchain-compensation 0.2.0

Automatic compensation middleware for LangChain agents with dependency-aware rollback

LangChain Compensation

Automatic compensation middleware for LangChain agents with LIFO (Last-In-First-Out) rollback capabilities. Inspired by the Saga pattern, this package provides automatic rollback of completed actions when a failure occurs in a multi-step agent workflow.

Features

• 🔄 Automatic Rollback: Automatically compensates completed actions when failures occur
• 📚 LIFO Order: Rolls back actions in reverse chronological order (like a transaction stack)
• 🎯 Simple API: Easy-to-use agent factory with compensation mapping
• 🔧 Flexible: Support for custom state mappers to extract compensation parameters
• 🧵 Thread-Safe: Built with thread safety in mind
• 📦 Zero Config: Works out of the box with sensible defaults

Installation

pip install langchain-compensation

Quick Start

from langchain_compensation import create_comp_agent
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_core.tools import tool

# Define your tools
@tool
def book_flight(destination: str) -> str:
    """Books a flight to the destination."""
    return f"flight_id_for_{destination}"

@tool
def cancel_flight(booking_id: str) -> str:
    """Cancels a flight booking."""
    return "Cancellation successful"

@tool
def book_hotel(location: str) -> str:
    """Books a hotel at the location."""
    if "fail" in location.lower():
        return "Error: Hotel booking failed!"
    return f"hotel_id_for_{location}"

@tool
def cancel_hotel(booking_id: str) -> str:
    """Cancels a hotel booking."""
    return "Cancellation successful"

# Create agent with compensation
model = ChatGoogleGenerativeAI(model="gemini-2.0-flash-exp")
agent = create_comp_agent(
    model=model,
    tools=[book_flight, cancel_flight, book_hotel, cancel_hotel],
    compensation_mapping={
        "book_flight": "cancel_flight",
        "book_hotel": "cancel_hotel"
    }
)

# Run the agent
result = agent.invoke({
    "messages": [("user", "Book a flight to London and a hotel in FailCity")]
})
# When hotel booking fails, the flight booking is automatically cancelled!

How It Works

1. Track Actions: Compensatable actions are tracked in the agent's state
2. Detect Failures: When a tool returns an error, the middleware detects it
3. Automatic Rollback: All completed compensatable actions are rolled back in LIFO order
4. Continue: The agent continues with the rolled-back state

Advanced Usage

Custom State Mappers

Sometimes you need custom logic to extract compensation parameters from the original result:

def extract_flight_id(result, original_params):
    """Extract booking ID from the result."""
    return {"booking_id": result["id"]}

agent = create_comp_agent(
    model=model,
    tools=tools,
    compensation_mapping={"book_flight": "cancel_flight"},
    state_mappers={"book_flight": extract_flight_id}
)

With Checkpointer

Use with LangGraph's checkpointing for persistent compensation logs:

from langgraph.checkpoint.memory import MemorySaver

memory = MemorySaver()
agent = create_comp_agent(
    model=model,
    tools=tools,
    compensation_mapping={"book_flight": "cancel_flight"},
    checkpointer=memory
)

API Reference

create_comp_agent

Creates a LangChain agent with automatic compensation capabilities.

Parameters:

• model (BaseChatModel | str | None): The LLM to use for the agent
• tools (Sequence[BaseTool | Callable]): List of tools available to the agent
• compensation_mapping (dict[str, str]): Maps tool names to their compensation tools
• state_mappers (dict[str, Callable] | None): Custom parameter extraction functions
• system_prompt (str | None): Additional system prompt for the agent
• checkpointer (Checkpointer | None): Optional checkpointer for state persistence
• Other standard LangChain agent parameters

Returns: CompiledStateGraph - A configured agent with compensation middleware

CompensationMiddleware

The core middleware that handles compensation logic.

Parameters:

• compensation_mapping (dict[str, str]): Maps tool names to compensation tools
• tools (list | None): Tools to cache for compensation execution
• state_mappers (dict[str, Callable] | None): Custom state mappers

Use Cases

• Travel Booking Systems: Cancel flights/hotels if any part of the trip fails
• E-commerce: Reverse inventory reservations if payment fails
• Multi-step Workflows: Undo completed steps when later steps fail
• Database Operations: Rollback related operations in distributed systems
• API Integrations: Clean up created resources when workflows fail

Requirements

• Python 3.9+
• langchain >= 0.3.0
• langchain-core >= 0.3.0
• langgraph >= 0.2.0

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

MIT License - see LICENSE file for details

Credits

Inspired by the Saga pattern and built on top of LangChain's middleware system.