langchain-langgraph

name	langchain-langgraph
description	Use when working with LangChain or LangGraph in this repo, especially for agent construction, middleware, state graphs, persistence, memory, retrieval, runtime context, or durable execution patterns.

Workflow

1. Open references/index.md first.
2. Pick the smallest matching reference file instead of loading everything.
3. Prefer the LangChain-side files for agent construction, prompts, tools, ToolRuntime, middleware, model routing, and structured output.
4. Prefer the LangGraph-side files for graph design, reducers, checkpoints, interrupts, subgraphs, streaming, and durable execution.
5. Prefer the memory and retrieval files for long-term memory, state/store separation, vector stores, loaders, splitters, embeddings, and RAG architecture.
6. Preserve the original source doc as the canonical reference for cross-checking; this skill is a bounded access layer, not a replacement source.

High-Value Repo Rules

• Prefer explicit LangGraph orchestration for complex workflows instead of burying a full agent loop inside a graph node.
• Reuse model and agent instances instead of rebuilding them inside each call path.
• Use structured output consistently for LLM and tool-facing boundaries when typed data matters.
• Keep async usage end-to-end across LangChain and LangGraph integrations.
• Preserve provider-specific message metadata when the model requires it across turns.
• Normalize persisted agent state when loading from checkpointers if schema or version drift is possible.
• Keep code examples close to the topic-specific references instead of collapsing them into abstract summaries.

LangChain LangGraph Reference Index

This skill is split into bounded reference files so opencode can load only the relevant material.

File Selection

• Use 01-langchain-overview.md for:
  • top-level defaults
  • anti-patterns
  • dedicated imports
  • agent reuse guidance
• Use 02-prompts-and-messages.md for:
  • prompt templates
  • message prompts
  • multimodal inputs
  • signature preservation
• Use 03-tools-and-toolruntime.md for:
  • @tool
  • args schemas
  • ToolRuntime reads and writes
  • tool streaming
  • tool errors
  • tool naming
• Use 04-model-selection-and-structured-output.md for:
  • model routing
  • model profiles
  • prompt caching
  • structured output
  • dynamic response formats
• Use 05-middleware-and-guardrails.md for:
  • node-style hooks
  • wrap-style hooks
  • state updates from middleware
  • guardrail composition
• Use 06-runtime-state-store-context.md for:
  • model context
  • tool context
  • lifecycle context
  • runtime context
  • state and store separation
• Use 07-langgraph-state-nodes-edges.md for:
  • StateGraph
  • reducers
  • nodes
  • edges
  • recursion limits
  • explicit orchestration
• Use 08-checkpointing-persistence-durability.md for:
  • checkpointers
  • thread IDs
  • checkpoint IDs
  • snapshots
  • serializers
  • durability modes
• Use 09-interrupts-hitl-resume.md for:
  • interrupts
  • resume semantics
  • idempotency around pauses
  • ordering hazards
• Use 10-subgraphs-and-streaming.md for:
  • subgraph patterns
  • per-thread vs stateless subgraphs
  • stream version="v2"
  • namespace behavior
• Use 11-multi-agent-patterns.md for:
  • subagents
  • handoffs
  • skills pattern
  • routers
  • custom workflows
• Use 12-memory.md for:
  • short-term memory
  • long-term memory
  • semantic, episodic, and procedural memory
  • file-memory conventions
• Use 13-retrieval-rag.md for:
  • loaders
  • splitters
  • embeddings
  • vector stores
  • retrievers
  • RAG architecture

Quick Reminders

• Do not create agent instances inside LangGraph nodes unless there is a narrow, well-justified reason.
• Keep model context, tool context, runtime context, state, and store conceptually separate.
• Use reducers for append-style state fields or they will be overwritten.
• Prefer the smallest reference file that answers the current question.