fix(langchain): prevent llmToolSelectorMiddleware from leaking into message stream

[Youngho Kim (JadenKim-dev)]

Summary

llmToolSelectorMiddleware internally calls structuredModel.invoke() to select relevant tools. When agent.stream() is used with streamMode: "messages", LangGraph injects a StreamMessagesHandler into config.callbacks and stores it in AsyncLocalStorage. Without an explicit config override, this handler is inherited by the internal invoke call, causing the tool selection response ({"tools":["..."]}) to appear as an assistant message in the UI stream.

Root cause

ensureConfig() in @langchain/core merges the explicit config with the implicit config from AsyncLocalStorage. If config.callbacks is undefined, the StreamMessagesHandler from the parent context is inherited. Passing callbacks: [] (an empty array) breaks this inheritance because a non-undefined value always overrides the implicit one.

LangSmith tracing is unaffected — it is injected via LangChainTracer.getTraceableRunTree() inside _configureSync, not through config.callbacks.

Fix

Build an explicit config using mergeConfigs:

const baseConfig: RunnableConfig = pickRunnableConfigKeys(request.runtime) ?? {};
const config = mergeConfigs(baseConfig, {
  metadata: { lc_source: "llmToolSelector" },
  callbacks: [],
});

• callbacks: [] prevents StreamMessagesHandler from being inherited via AsyncLocalStorage
• pickRunnableConfigKeys inherits the parent config (tags, metadata, configurable, etc.) from runtime
• lc_source: "llmToolSelector" tags the call for observability, consistent with summarizationMiddleware

Fixes #10042

Test plan

• Unit tests: yarn test src/agents/middleware/tests/llmToolSelector.test.ts
• Integration test verifying no stream leak: OPENAI_API_KEY=... yarn vitest run --mode int src/agents/middleware/tests/llmToolSelector.int.test.ts

[changeset-bot]

🦋 Changeset detected

Latest commit: ff99621

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
langchain	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[Christian Bromann (christian-bromann)]

We shouldn't need a real model for this and instead just use a mock as documented in https://docs.langchain.com/oss/javascript/langchain/test/unit-testing

[Youngho Kim (JadenKim-dev)]

Christian Bromann (@christian-bromann)
Replaced the integration test with FakeListChatModel so it no longer needs a real API call.

[Christian Bromann (christian-bromann)]

LGTM 👍

Thanks for the update!