LangChain.js 1.6.5 adds decision models and router middleware, and now requires Zod 4.6.6 or newer (November 11, 2026)

What changed

On November 11, 2026 LangChain.js released langchain 1.6.5 and @langchain/core 1.4.0. This is the JavaScript side of the decision model work that LangChain 1.5 brought to Python on the same day.

@langchain/core 1.4.0 adds a new entry point, @langchain/core/decisions. It holds BaseDecisionModel, a Runnable that answers typed Noul, Choice, Score and Box questions about text, messages or JSON state. The answers from invoke are typed from the questions you ask, including Choice labels. A provider response that leaves a question unanswered, or answers it with the wrong type or label, is rejected.

langchain 1.6.5 adds two middleware built on decision models:

• modelRouterMiddleware picks the agent's model for each run from a decision model's Choice answer. It classifies the latest human message once per run and stores the answer in state as modelRoute.
• autoModeMiddleware blocks listed tool calls that a decision model judges risky. A call at or above 0.5 returns an error ToolMessage without running the tool. Tools you do not list are not classified.

LangChain's JavaScript changelog marks both middleware as beta.

Two things can break an existing project:

• Both packages now require Zod 4.6.6 or newer, to exclude versions affected by GHSA-mrhx-c7xc-m3w5. On npm, langchain 1.6.5 and @langchain/core 1.4.0 list zod as ^4.6.6. Zod v3 schemas remain supported through the bundled zod/v3 entry point.
• langchain 1.6.5 needs @langchain/core ^1.4.0 as a peer dependency.

@langchain/typesafe 0.1.0, released the same day, has breaking changes. TypeSafeClassifier now extends BaseDecisionModel, and questions move from the constructor to invoke: new TypeSafeClassifier().invoke({ state, questions }) replaces new TypeSafeClassifier({ questions }).invoke(state). Token usage moves from to , and every question now requires . The and in are deprecated and will be removed in 0.2.0.