claude-scaffold: Skill for Claude Code

LangGraph Patterns

When to Load This Skill

Load when working with: LangGraph state machines, agent nodes, tool definitions, checkpointers, human-in-the-loop interrupts, multi-agent coordination.

Current Version

LangGraph >=0.2.0 (langgraph-checkpoint for persistence). Always pin exact version in pyproject.toml.

Core Concepts

LangGraph models agent workflows as directed graphs:

• State: typed dict passed between nodes (the single source of truth)
• Nodes: Python async functions that receive and return state updates
• Edges: routing logic — conditional or unconditional
• Checkpointer: persistence layer for long-running agents (SQLite locally, PostgreSQL in production)

Standard Project Structure

src/{project_name}/
├── agents/
│   ├── __init__.py
│   ├── graph.py           # graph assembly
│   ├── state.py           # TypedDict state definition
│   ├── nodes/
│   │   ├── __init__.py
│   │   ├── analyst.py
│   │   └── writer.py
│   └── tools/
│       ├── __init__.py
│       └── search.py

State Definition Pattern

from typing import Annotated
from typing_extensions import TypedDict
import operator


class AgentState(TypedDict):
    messages: Annotated[list[dict], operator.add]
    user_input: str
    retrieved_context: list[str]
    final_answer: str | None
    error: str | None
    iteration_count: int

Use Annotated[list, operator.add] for lists that nodes append to. Use plain types for values that nodes replace entirely.

Node Pattern

from langchain_core.messages import AIMessage

from src.project_name.agents.state import AgentState
from src.project_name.adapters.llm.claude_adapter import ClaudeAdapter


async def analyst_node(state: AgentState) -> dict:
    adapter = ClaudeAdapter()
    response = await adapter.invoke(
        system="You are a precise analyst. Answer based only on retrieved context.",
        messages=state["messages"],
        context=state["retrieved_context"],
    )
    return {
        "messages": [AIMessage(content=response)],
        "final_answer": response,
    }