langchain-ai/langchainjs langchain==1.0.0-alpha.4

on GitHub

Release range: langchain==0.3.30 → langchain==1.0.0-alpha.1
Compare: langchain==0.3.30...langchain==1.0.0-alpha.1

💥 Breaking changes

• Require Node.js ≥ 20
  Langchain packages now require Node 20+ (.nvmrc is on Node 24 for development). Update your engines and CI accordingly. (GitHub)

  // package.json
  { "engines": { "node": ">=20" } }

• Removed legacy subpath exports (cleaner surface area)

  • langchain/schema/prompt_template → use langchain/prompts. (GitHub, api.js.langchain.com)

    - import { PromptTemplate } from "langchain/schema/prompt_template";
    + import { PromptTemplate } from "langchain/prompts";

  • langchain/schema/query_constructor → use langchain/chains/query_constructor. (GitHub, v03.api.js.langchain.com, Langchain)

    - import type { AttributeInfo } from "langchain/schema/query_constructor";
    + import type { AttributeInfo } from "langchain/chains/query_constructor";

  • langchain/runnables/remote export removed. If you used this entrypoint, migrate off this subpath (no 1:1 replacement export in langchain package). (GitHub)

  • langchain/smith export path removed from langchain. Use the separate langsmith package and opt-in (see below). (GitHub)

• Callbacks removed from langchain
  The ./callbacks entrypoint is gone. Prefer LCEL observability (runnables), agent/tool messages, or provider-native hooks instead. (GitHub)

  - import { ... } from "langchain/callbacks";
  // Replace with runnable config / tracing via LangSmith (optional) or agent nodes.

✨ Features

• langchain now exports most important user primitives
  The langchain package now exports the most important user primitives, including tool, HumanMessage, AIMessage, SystemMessage, ToolNode, createAgent, and more.

  import { tool, HumanMessage, createAgent } from "langchain";

• createAgent is now part of core langchain
  The React-style agent lives under the langchain project (with examples + tests). Import directly from langchain. (GitHub)

  import { createAgent, HumanMessage, tool } from "langchain";
  import { z } from "zod";
  
  const getWeather = tool({
    name: "getWeather",
    description: "Get current weather by city",
    schema: z.object({ city: z.string() }),
    func: async ({ city }) => `Sunny in ${city}`
  });
  
  const agent = await createAgent({
    // New: pass a model by name (string) or a model instance
    model: "claude-3-5-sonnet-20240620",
    tools: [getWeather],
    responseFormat: z.object({ answer: z.string() }),
  });
  
  const res = await agent.invoke(new HumanMessage("Weather in SF?"));
  console.log(res);

  Now you can pass in a “model as string” option to define the model to use. (GitHub)

• ToolNode exported from langchain
  Build agent graphs that execute tools as a node. This makes tool execution composable within LCEL/graph flows.

  import { ToolNode, tool, HumanMessage } from "langchain";
  
  const fail = tool({
    name: "alwaysFails",
    description: "Demo failing tool",
    schema: {},
    func: async () => { throw new Error("Boom"); }
  });
  
  const node = new ToolNode([fail]);
  const out = await node.invoke({ messages: [new HumanMessage("Try it")] });

• Default tool-error handling in agents
  When handleToolErrors is true (default), tool exceptions are caught and converted into a ToolMessage so the agent can recover. Set it to false to surface raw errors for strict workflows. (GitHub)

  import { ToolNode } from "langchain/agents";
  
  // default: handleToolErrors: true → returns a ToolMessage with error text
  const forgiving = new ToolNode([/* tools */]);
  
  // strict: throw on tool error
  const strict = new ToolNode([/* tools */], { handleToolErrors: false });

🐛 Fixes

• langsmith made optional
  LangSmith is no longer a hard dependency; the package is loaded dynamically only when you use features that need it. If you rely on eval/tracing/hub, add it explicitly:
  pnpm add langsmith (or yarn/npm). (GitHub)

• Removed js-tiktoken dependency
  The package was dropped from deps; related types are sourced via the relevant provider packages (e.g. @langchain/openai). If you had direct imports, adjust accordingly. (GitHub)

🧹 Chores / Internal

• New tsdown-based build system (faster, simpler internal builds). (GitHub)
• Repo migrated to pnpm (lockfile + scripts updated). (GitHub)
• Rearranged packages and updated CI/workflows. (GitHub)
• Removed experimental and unused files/exports (./indexes, ./document_transformers/openai_functions, agents cleanup). User-visible API impact covered above via export removals; the rest is housekeeping to keep the surface tight. (GitHub)
• Environment tests rewritten in TS and expanded to verify export maps across bundlers/runtimes. (GitHub)

🧭 Migration quick-guide

1. Node version

   • Set engines.node to >=20 and update CI runners. (GitHub)

2. Imports / subpaths

   • schema/prompt_template → langchain/prompts (e.g., PromptTemplate). (GitHub, api.js.langchain.com)
   • schema/query_constructor → langchain/chains/query_constructor (IR types, translators, etc.). (GitHub, v03.api.js.langchain.com, Langchain)
   • Remove langchain/runnables/remote usage; no longer exported. (GitHub)
   • Remove langchain/smith import; install and import from langsmith if needed. (GitHub)

3. Agents

   • Import prebuilt agents from langchain directly, e.g. import { createAgent } from "langchain".
   • You can pass model as a model name string to createAgent. (GitHub)

     const agent = createAgent({
       model: "openai:gpt-4o-mini",
       tools: [/* tools */],
       responseFormat: z.object({ answer: z.string() }),
     });

   • Use ToolNode to encapsulate tool execution in graphs.
   • Decide on error policy: default soft-handling vs strict throws in ToolNode via handleToolErrors. (GitHub)

4. LangSmith & tokenization

   • Add langsmith explicitly if you use tracing/eval/hub. (GitHub)
   • If you previously relied on js-tiktoken, remove it from your app unless you need it directly. (GitHub)