_formatForTracing only converts the first base64 content block per message #11291

Checked other resources

• This is a bug, not a usage question. For questions, please use the LangChain Forum (https://forum.langchain.com/).
• I added a very descriptive title to this issue.
• I searched the LangChain.js documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.
• I am sure that this is a bug in LangChain.js rather than my code.
• The bug is not resolved by updating to the latest stable version of LangChain (or the specific integration package).

Example Code

import { BaseCallbackHandler } from "@langchain/core/callbacks/base";
import type { Serialized } from "@langchain/core/load/serializable";
import { HumanMessage } from "@langchain/core/messages";
import type { BaseMessage } from "@langchain/core/messages";
import { FakeListChatModel } from "@langchain/core/utils/testing";

class CaptureInputHandler extends BaseCallbackHandler {
  name = "CaptureInputHandler";
  messages: BaseMessage[] | undefined;

  async handleChatModelStart(
    _llm: Serialized,
    messages: BaseMessage[][],
  ): Promise<void> {
    this.messages = messages[0];
  }
}

// Shape still shown in the JS multimodal docs:
// https://docs.langchain.com/oss/javascript/langchain/messages#multimodal
const message = new HumanMessage({
  content: [
    { type: "text", text: "two PDFs" },
    {
      type: "file",
      source_type: "base64",
      mime_type: "application/pdf",
      data: "JVBERi0xLjQK",
    },
    {
      type: "file",
      source_type: "base64",
      mime_type: "application/pdf",
      data: "JVBERi0xLjUK",
    },
  ],
});

const handler = new CaptureInputHandler();
await new FakeListChatModel({ responses: ["ok"] }).invoke([message], {
  callbacks: [handler],
});

console.log(JSON.stringify(handler.messages?.[0]?.content, null, 2));

Error Message and Stack Trace (if applicable)

No response

Description

• I'm trying to trace chat model inputs that contain several base64 file attachments in one HumanMessage using the data content block shape shown in the TypeScript > Messages > Multimodal docs (type + source_type + data + mime_type). Observability tools (e.g., Langfuse) rely on _formatForTracing rewriting those blocks into OpenAI image_url data URIs so media can be detected and offloaded.
• I expect every matching base64/URL block in the message to be converted the same way before handleChatModelStart fires (tracing copy only — the provider still receives the original message).
• Instead, only the first matching block is converted. Later blocks are left as authored file/image content, so their base64 stays inline in the trace.

Actual output

[
  {
    "type": "text",
    "text": "two PDFs"
  },
  {
    "type": "image_url",
    "image_url": {
      "url": "data:application/pdf;base64,JVBERi0xLjQK"
    }
  },
  {
    "type": "file",
    "source_type": "base64",
    "mime_type": "application/pdf",
    "data": "JVBERi0xLjUK"
  }
]

The second attachment is never rewritten, so its base64 stays inline in the trace.

Expected output

[
  {
    "type": "text",
    "text": "two PDFs"
  },
  {
    "type": "image_url",
    "image_url": {
      "url": "data:application/pdf;base64,JVBERi0xLjQK"
    }
  },
  {
    "type": "image_url",
    "image_url": {
      "url": "data:application/pdf;base64,JVBERi0xLjUK"
    }
  }
]

Root cause

In _formatForTracing (https://github.com/langchain-ai/langchainjs/blob/main/libs/langchain-core/src/language_models/chat_models.ts#L171), the clone guard if (messageToTrace === message) also gates the conversion. After the first rewrite, messageToTrace !== message, so subsequent blocks are skipped.

Docs / shape note

The conversion path only understands the deprecated isBase64ContentBlock / isURLContentBlock helpers (source_type). Standard multimodal blocks are not converted at all:

So today the only way to get attachments traced as data URIs is the deprecated shape — and even then only one attachment per message is rewritten.

Proposed fix

Clone once, then convert every matching block. We’ve verified this locally with a pnpm patch on @langchain/core@1.2.3.

System Info

@langchain/core: 1.2.3 (also present on main as of this report)
langchain: 1.5.3
Platform: macOS
Node: 24.18.0
pnpm: 11.15.1