Bug: merge_dicts concatenates ls_provider across streaming chunks, inflating response_metadata

[GiovaneIwamoto]

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

• ChatBedrock: response_metadata["model_provider"] disagrees with _get_ls_params()["ls_provider"] langchain-aws#946 — ls_provider vs model_provider mismatch
• fix(aws): Add ls_provider to message response_metadata langchain-aws#981 — PR that adds ls_provider to streaming chunks
• merge_dicts() concatenates parallel tool calls during streaming #34807 — similar merge_dicts concatenation issue with tool call fields

Reproduction Steps / Example Code (Python)

from langchain_core.utils._merge import merge_dicts

# Simulate two streaming chunks from ChatBedrockConverse (langchain-aws>=1.4.5).
# Each chunk carries ls_provider="amazon_bedrock" in response_metadata.
left = {"ls_provider": "amazon_bedrock", "model_provider": "bedrock_converse"}
right = {"ls_provider": "amazon_bedrock", "model_provider": "bedrock_converse"}
result = merge_dicts(left, right)

print("model_provider:", result["model_provider"])  # "bedrock_converse" (correct)
print("ls_provider:", result["ls_provider"])         # "amazon_bedrockamazon_bedrock" (BUG)

# After 20 streaming chunks (typical response):
metadata = {"ls_provider": "amazon_bedrock"}
for _ in range(19):
    metadata = merge_dicts(metadata, {"ls_provider": "amazon_bedrock"})
print(f"After 20 chunks: {len(metadata['ls_provider'])} chars")  # 280 instead of 14

Error Message and Stack Trace (if applicable)

No explicit error is raised. The reproduction above shows that ls_provider grows
to N * len("amazon_bedrock") after N streaming chunks. In our production
environment (LangGraph agent, 58+ graph steps, langchain-aws==1.4.5), we observed
massively inflated ls_provider fields across all AIMessages in conversation history,
contributing to request payload bloat.

Description

langchain-aws v1.4.5 (PR langchain-ai/langchain-aws#981) adds response_metadata["ls_provider"] = "amazon_bedrock" to each streaming chunk in ChatBedrockConverse._stream(). This is needed for SummarizationMiddleware to identify the provider from message metadata.

However, merge_dicts() concatenates string values when merging chunks. The skip-concatenation guard on line 59 of _merge.py protects model_provider (also set on every chunk) but does not protect ls_provider:

right_k in {"id", "output_version", "model_provider"}  # ls_provider missing

After merging ~20 chunks, ls_provider becomes "amazon_bedrock" repeated 20 times (280 chars). Every AIMessage in conversation history carries this inflated metadata.

Proposed fix: add "ls_provider" to the existing guard, matching the pattern already established for "model_provider". I have a fix ready with tests if assigned.

System Info

System Information

OS: Darwin
OS Version: Darwin Kernel Version 25.4.0: Thu Mar 19 19:33:09 PDT 2026; root:xnu-12377.101.15~1/RELEASE_ARM64_T8112
Python Version: 3.14.3 (main, Feb 3 2026, 15:32:20) [Clang 17.0.0 (clang-1700.6.3.2)]

Package Information

langchain_core: 1.3.2
langsmith: 0.7.36
langchain_protocol: 0.0.11

Optional packages not installed

deepagents
deepagents-cli

Other Dependencies

httpx: 0.28.1
jsonpatch: 1.33
orjson: 3.11.8
packaging: 26.1
pydantic: 2.13.3
pytest: 9.0.3
pyyaml: 6.0.3
requests: 2.33.1
requests-toolbelt: 1.0.0
tenacity: 9.1.4
typing-extensions: 4.15.0
uuid-utils: 0.14.1
xxhash: 3.6.0
zstandard: 0.25.0

[gitbalaji]

Hi, I'd like to work on this. Could a maintainer please assign it to me?

[GiovaneIwamoto]

I already have a fix ready with passing tests on my fork. I'd like to be assigned to submit the PR.

The fix adds ls_provider to the existing skip-concatenation guard in merge_dicts(), matching the pattern already established for model_provider:

GiovaneIwamoto@1b525a88ed

[QEEK-AI]

QEEK.AI Agent :

"This is a valid bug report. The analysis provided in the issue is correct: merge_dicts in langchain-core currently lacks a guard for the ls_provider key, leading to string concatenation during the merging of streaming chunks".

Technical Analysis

The issue resides in libs/core/langchain_core/utils/_merge.py. The merge_dicts function handles string values by concatenating them unless the key is explicitly excluded in a "skip-concatenation" guard.

Currently, the guard (around line 60) only includes id, output_version, and model_provider:

# libs/core/langchain_core/utils/_merge.py

                if (right_k == "index" and merged[right_k].startswith("lc_")) or (
                    right_k in {"id", "output_version", "model_provider"}
                    and merged[right_k] == right_v
                ):
                    continue
                merged[right_k] += right_v

When ls_provider (e.g., "amazon_bedrock") is present in multiple chunks with the same value, it falls through to merged[right_k] += right_v. After 20 chunks, this results in the string being repeated 20 times, significantly inflating the response_metadata and causing payload bloat in downstream components like LangGraph.

Proposed Fix

The proposed fix is to add "ls_provider" to the set of protected keys. This ensures that if the value is identical across chunks, the merge logic will continue (skip) instead of concatenating.

                if (right_k == "index" and merged[right_k].startswith("lc_")) or (
                    right_k in {"id", "output_version", "model_provider", "ls_provider"}
                    and merged[right_k] == right_v
                ):
                    continue

This change is consistent with how model_provider is already handled and correctly addresses the regression introduced by adding ls_provider to streaming chunks in provider packages (like langchain-aws).

[jasiecky]

Hey, I would like to work on this issue;) I've already opened a PR.