May 19, 2026

Aggregating Anthropic Streams in LangChain.js

This happens to be my first contribution to LangChain.

I chose the TypeScript path because I want to stay deeply rooted in the web ecosystem while engineering with AI tools, learning about agents, and understanding orchestration systems at a lower level.

The problem here was to prevent an interpretive failure introduced by a newer update that enforced stricter validation on streaming data frames.

The data stream is processed through @langchain/core, which acts as the foundational orchestration layer for the broader LangChain ecosystem inside the local Node.js runtime.

In v1.1.46, a stricter security/type guard was introduced.

The initial stream frame emits JSON with:

{
  "type": "tool_use"
}

Subsequent delta frames emit:

{
  "type": "input_json_delta"
}

The issue originates in getMergeableTypeBase().

The function normalizes frame types by truncating the _delta suffix. As a result:

input_json_delta → input_json

That normalized value is then compared against "tool_use".

Because of the stricter validation introduced in the newer pull request, this mismatch gets flagged and throws an OUTPUT_PARSING_FAILURE.

My fix was to update getMergeableTypeBase() so that both:

input_json
input_json_delta

correctly map to:

"tool_use"

I also integrated aggregation tests for this logic.

Initially, I added the tests inside the @langchain/anthropic provider package.

After review feedback from Bromann, who works on agentic workflows at LangChain, I moved the tests into the LangChain core workspace instead.

I also handled formatting and styling updates to ensure the CI formatting checks passed successfully.