langchain - 💡(How to fix) Fix core: AIMessageChunk validator escapes non-ASCII in tool_call_chunks[].args

Submission checklist

• This is a bug, not a usage question.
• I added a clear and descriptive title that summarizes this issue.
• 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 rather than my code.
• The bug is not resolved by updating to the latest stable version of LangChain (or the specific integration package).
• This is not related to the langchain-community package.
• I posted a self-contained, minimal, reproducible example. A maintainer can copy it and run it AS IS.

Package (Required)

• langchain
• langchain-openai
• langchain-anthropic
• langchain-classic
• langchain-core
• langchain-model-profiles
• langchain-tests
• langchain-text-splitters
• langchain-chroma
• langchain-deepseek
• langchain-exa
• langchain-fireworks
• langchain-groq
• langchain-huggingface
• langchain-mistralai
• langchain-nomic
• langchain-ollama
• langchain-openrouter
• langchain-perplexity
• langchain-qdrant
• langchain-xai
• Other / not sure / general

Related Issues / PRs

Same class of bug as langchain-ai/langchain-google#1789 (just-shipped fix in langchain-google-genai's _parse_response_candidate). Issue #34005 ("PydanticOutputParser formats error with ensure_ascii=True") was a different site closed earlier. The convention is already established elsewhere in this repo — e.g. libs/core/langchain_core/messages/utils.py:1810 already passes ensure_ascii=False.

Reproduction Steps / Example Code (Python)

from langchain_core.messages import AIMessageChunk

chunk = AIMessageChunk(
    content="",
    tool_calls=[
        {
            "name": "echo",
            "args": {"text": "你好", "emoji": "🚀"},
            "id": "call_1",
            "type": "tool_call",
        }
    ],
)

print("tool_calls[0]['args']:        ", chunk.tool_calls[0]["args"])
print("tool_call_chunks[0]['args']:  ", repr(chunk.tool_call_chunks[0]["args"]))

assert "你好" in chunk.tool_call_chunks[0]["args"], "CJK was escaped to \\uXXXX"

Error Message and Stack Trace (if applicable)

tool_calls[0]['args']:         {'text': '你好', 'emoji': '🚀'}
tool_call_chunks[0]['args']:   '{"text": "\\u4f60\\u597d", "emoji": "\\ud83d\\ude80"}'
Traceback (most recent call last):
  File "...", line N, in <module>
    assert "你好" in chunk.tool_call_chunks[0]["args"], "CJK was escaped to \\uXXXX"
AssertionError: CJK was escaped to \uXXXX

Description

Buggy location: libs/core/langchain_core/messages/ai.py:509

# libs/core/langchain_core/messages/ai.py — AIMessageChunk.init_tool_calls
if not self.tool_call_chunks:
    if self.tool_calls:
        self.tool_call_chunks = [
            create_tool_call_chunk(
                name=tc["name"],
                args=json.dumps(tc["args"]),                # ← bug: defaults to ensure_ascii=True
                id=tc["id"],
                index=None,
            )
            for tc in self.tool_calls
        ]

AIMessageChunk.init_tool_calls is a @model_validator(mode="after") that back-fills tool_call_chunks from tool_calls when only the latter is supplied at construction. It uses json.dumps(...) without ensure_ascii=False, so every non-ASCII character in tool-call arguments — CJK, emoji, accented Latin — is escaped to \uXXXX in the resulting chunk's args string.

Why it matters

1. Inconsistency on the same object: tool_calls[i]["args"] is a clean dict ({'text': '你好'}), but tool_call_chunks[i]["args"] is an escaped string ('{"text": "\\u4f60\\u597d"}'). Consumers see different content depending on which field they read.
2. Persistence layers (DB JSON columns, log aggregators, langgraph checkpointers, langsmith traces) round-trip through tool_call_chunks, so foreign-language tool args land in storage as escape sequences and are unreadable on inspection.
3. Cross-package convention already exists: libs/core/langchain_core/messages/utils.py:1810 passes ensure_ascii=False, and langchain-openai follows the same convention in langchain_openai/chat_models/base.py. This site is an outlier.

Proposed fix (one keyword)

args=json.dumps(tc["args"], ensure_ascii=False),

Related sites with the same anti-pattern in this repo (out of scope for this issue, can be addressed in follow-up PRs):

• libs/core/langchain_core/messages/utils.py:1880 — Bedrock/Converse "text" block
• libs/partners/openai/langchain_openai/chat_models/_compat.py:358, 436
• libs/partners/anthropic/langchain_anthropic/chat_models.py:1519
• libs/partners/anthropic/langchain_anthropic/experimental.py:130

System Info

> Python Version:  3.12
> langchain-core: 1.4.0 (master @ 33875fdd...)

If acceptable, I have the fix and a regression test ready locally — could you please assign this to me so I can open a PR? The fix is a one-keyword change in ai.py plus a test_messages.py regression test asserting that tool_call_chunks[0]["args"] preserves CJK / emoji from tool_calls.

Disclaimer: this report was prepared with the assistance of an AI agent (Claude Code); the bug was independently reproduced and reviewed by me before submission.