LangChain Python Skill Pack (v2.0)

Pain-first Claude Code skills for LangChain 1.0 and LangGraph 1.0 in Python. Every skill opens with a specific production failure mode — not capability prose.

Why This Pack Exists

LangChain 1.0 and LangGraph 1.0 shipped October 2025 with a new middleware model, typed content blocks (text/tool_use/image), stable checkpointing, first-class human-in-the-loop interrupts, three streaming modes, and native OpenTelemetry export. Token accounting, structured output, agent control flow, and memory all behave differently than they did in the 0.2 / 0.3 era.

This pack replaces the legacy langchain-pack with pain-first, Python-native skills. Every skill opens with a specific failure mode — a real exception, a hardcoded threshold, a version-specific regression — not capability prose:

ChatAnthropic.stream() blocks llmOutput token counts until the stream completes. Live cost dashboards built on that field lag by stream_duration seconds — sometimes by 20+ seconds on long responses. Use astream_events(version="v2") or a callback handler to read tokens incrementally.

See docs/pain-catalog.md for the full 68-entry catalog of LangChain 1.0 pain points that every skill in this pack anchors to.

Installation

/plugin install langchain-py-pack@claude-code-plugins-plus

TypeScript Counterpart

For LangChain.js + @langchain/langgraph in Node 22+ / Vercel / Cloud Run: langchain-ts-pack (Epic B — not yet shipped). Same 34-skill taxonomy, JS-native.

What Is Shipped Today

| Skill | What it covers | |-------|-------------| | langchain-model-inference | ChatAnthropic / ChatOpenAI / ChatGoogleGenerativeAI init, content-block iteration, streaming token accounting, structured-output method decision tree — with 4 deep references | | langchain-embeddings-search | FAISS vs Pinecone flipped score semantics, dim-mismatch prevention, language-aware chunking, hybrid BM25 + vector, reranker filter-by-rank — with 3 deep references |

Both skills are gold-standard quality (≥200 lines each, 3-4 references per skill, every example pinned to langchain-core 1.0.x, A-grade at the 100-point rubric).

Roadmap — 32 More Skills Landing Across Four Epics

Each epic is one PR. Every skill follows the same gold-standard quality bar as the two shipped above: concrete pain in the Overview, ≥2 error codes named, 2-4 references, decision trees / comparison tables where applicable, and every code block pinned to LangChain 1.0.x.

Epic A2 — Getting Started + Core Workflows (S01-S08 minus the two shipped)

| Code | Skill | Description | |---|---|---| | S01 | langchain-install-auth | Install langchain, langchain-core, provider packages; env var management; verify connectivity | | S02 | langchain-hello-world | Minimal ChatAnthropic chain with with_structured_output(), streaming, and token counting | | S04 | langchain-common-errors | 12+ real error codes with exact fixes (OutputParserException, RateLimitError, GraphRecursionError, agent-loop timeouts) | | S05 | langchain-sdk-patterns | RunnableSequence, .with_fallbacks(), .batch(), .abatch(), retries, concurrency caps | | S06 | langchain-core-workflow | RunnableParallel, RunnableBranch, RunnablePassthrough.assign(), RAG composition | | S08 | langchain-data-handling | Document loaders, RecursiveCharacterTextSplitter, semantic vs fixed chunking |

Epic A3 — Operations + Pro (S09-P20)

| Code | Skill | Description | |---|---|---| | S09 | langchain-observability | LangSmith zero-code tracing, OTEL native export, custom metric callbacks | | S10 | langchain-debug-bundle | astream_events(version="v2"), trace callbacks, LangSmith export, diagnostic dump | | S11 | langchain-incident-runbook | LLM-specific SLOs, p95 latency triage, provider outage runbook, cost-overrun response | | S12 | langchain-prod-checklist | 30-item go-live checklist with concrete thresholds | | S13 | langchain-ci-integration | GitHub Actions, FakeListChatModel, test gates, dry-run validators | | S14 | langchain-deploy-integration | LangServe, Cloud Run, Vercel Python runtime, secret management | | P15 | langchain-performance-tuning | Streaming modes, batch concurrency, semantic caching, RedisChatMessageHistory | | P16 | langchain-cost-tuning | Real token accounting, model tiering, cache hit rates, per-tenant budgets | | P17 | langchain-rate-limits | asyncio.Semaphore, token-bucket, exponential backoff, provider-specific limits | | P18 | langchain-security-basics | Prompt injection defenses, tool allowlisting, PII redaction, output validation | | P19 | langchain-enterprise-rbac | Tenant isolation, per-tenant rate limits, role-scoped retrievers, audit logs | | P20 | langchain-multi-env-setup | Pydantic Settings env validation, dev/staging/prod isolation, secret backends |

Epic A4 — Flagship + LangGraph v1.0 (F21-L34)

| Code | Skill | Description | |---|---|---| | F21 | langchain-reference-architecture | Layered design, LLM factory, chain registry, DI, tenant-scoped vector stores | | F22 | langchain-webhooks-events | Async callback handlers, SSE streaming, WebSocket, background dispatch | | F23 | langchain-local-dev-loop | pytest, FakeListChatModel, VCR fixtures, integration-test gating | | F24 | langchain-upgrade-migration | 0.2 → 0.3 → 1.0 migration with named breaking changes, codemod hints | | L25 | langchain-langgraph-basics | StateGraph, typed TypedDict state, nodes, edges, compile(), recursion limits | | L26 | langchain-langgraph-agents | create_react_agent, prebuilt tool-calling, tools_condition, agent loop caps | | L27 | langchain-langgraph-checkpointing | MemorySaver, PostgresSaver, thread_id semantics, time-travel | | L28 | langchain-langgraph-human-in-loop | interrupt_before, interrupt_after, Command(resume=...), approval flows | | L29 | langchain-langgraph-streaming | stream_mode "messages" vs "updates" vs "values", token-level streaming | | L30 | langchain-langgraph-subgraphs | Composing graphs, nested agent teams, shared state, subgraph boundaries | | L31 | langchain-middleware-patterns | 1.0 middleware model, PII redaction, caching, retry middleware, ordering rules | | L32 | langchain-deep-agents | Deep Agents: planner + subagents + virtual filesystem + reflection loop | | L33 | langchain-otel-observability | Native OTEL export, Jaeger/Honeycomb, LLM-specific SLO dashboards | | L34 | langchain-content-blocks | Typed AIMessage.content, Claude tool_use iteration quirks |

Quick Start

1. Install the pack

/plugin install langchain-py-pack@claude-code-plugins-plus

2. Install LangChain 1.0 + LangGraph 1.0

python -m venv .venv && source .venv/bin/activate

pip install "langchain>=1.0,<2.0" "langchain-core>=1.0,<2.0" \
            "langchain-anthropic>=1.0,<2.0" \
            "langgraph>=1.0,<2.0"

3. A minimal agent with memory

from langchain_anthropic import ChatAnthropic
from langgraph.prebuilt import create_react_agent
from langgraph.checkpoint.memory import MemorySaver

llm = ChatAnthropic(model="claude-sonnet-4-6", temperature=0)

def add(a: int, b: int) -> int:
    """Add two integers."""
    return a + b

agent = create_react_agent(
    model=llm,
    tools=[add],
    checkpointer=MemorySaver(),
)

config = {"configurable": {"thread_id": "demo-1"}}
result = agent.invoke(
    {"messages": [("user", "What is 17 + 25?")]},
    config=config,
)
print(result["messages"][-1].content)

Key LangChain 1.0 / LangGraph 1.0 Links

• LangChain 1.0 release notes
• LangChain Python docs
• LangGraph overview
• LangGraph streaming modes
• astream_events v2
• Checkpointing and persistence
• Human-in-the-loop patterns
• LangSmith
• State of Agent Engineering 2026

Version Baseline

Every skill in this pack is pinned to langchain-core 1.0.x / langgraph 1.0.x. If you are on 0.2.x or 0.3.x, wait for langchain-upgrade-migration (Epic A4) or consult the legacy langchain-pack in the meantime.

License

MIT