ChatAnthropic streaming drops thinking field from empty thinking blocks (signature_delta-only), replay fails with 400 thinking.thinking: Field required #38639

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

#36288

Reproduction Steps / Example Code (Python)

# Deterministic reproduction — no API key / network needed.
# Simulates the exact event sequence Anthropic streams for an adaptive-thinking
# block whose thinking text is empty: content_block_start (thinking="") followed
# by signature_delta only, with NO thinking_delta in between.
from anthropic.types import (
    RawContentBlockDeltaEvent,
    RawContentBlockStartEvent,
    SignatureDelta,
    ThinkingBlock,
)
from langchain_anthropic.chat_models import ChatAnthropic
from langchain_core.messages import AIMessage, HumanMessage

llm = ChatAnthropic(model="claude-opus-4-6", api_key="test")

events = [
    RawContentBlockStartEvent(
        type="content_block_start",
        index=0,
        content_block=ThinkingBlock(type="thinking", thinking="", signature=""),
    ),
    RawContentBlockDeltaEvent(
        type="content_block_delta",
        index=0,
        delta=SignatureDelta(type="signature_delta", signature="EqQBCkYIBRgCIkDaq3rk..."),
    ),
]

aggregated = None
block_start_event = None
for event in events:
    chunk, block_start_event = llm._make_message_chunk_from_anthropic_event(
        event,
        stream_usage=True,
        coerce_content_to_string=False,
        block_start_event=block_start_event,
    )
    if chunk is not None:
        aggregated = chunk if aggregated is None else aggregated + chunk

print(aggregated.content)
# [{'signature': 'EqQBCkYIBRgCIkDaq3rk...', 'type': 'thinking', 'index': 0}]
#  -> the `thinking` key is MISSING

assert "thinking" not in aggregated.content[0]  # bug: passes

# Replaying this message on the next model call (e.g. agent loop after a tool
# call) produces an invalid payload that the Anthropic API rejects with 400:
payload = llm._get_request_payload(
    [HumanMessage("hi"), AIMessage(content=aggregated.content), HumanMessage("continue")]
)
print(payload["messages"][1]["content"][0])
# {'signature': 'EqQBCkYIBRgCIkDaq3rk...', 'type': 'thinking'}  -> API returns 400

Error Message and Stack Trace (if applicable)

anthropic.BadRequestError: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'messages.1.content.0.thinking.thinking: Field required'}, 'request_id': 'req_011Ccew3oAKvavEEJnTroFjZ'}

Description

• With adaptive thinking (e.g. claude-opus-4-6), the API can return a thinking block whose thinking text is empty — typically on the assistant turn that follows a tool result. In streaming, such a block arrives as content_block_start (with thinking="") followed only by signature_delta events; no thinking_delta is ever emitted.
• I expect the aggregated AIMessage to contain a thinking block with thinking: "" plus the signature, so the message can be replayed on the next model call (required for tool-calling agent loops, where thinking blocks must be sent back).
• Instead, the aggregated block is {'type': 'thinking', 'signature': '...'} with no thinking key, and the next request fails with 400 - messages.N.content.M.thinking.thinking: Field required. In an agent loop this makes every tool-calling conversation break on the second model call, and the corrupted message is persisted in checkpointers.

Root cause in libs/partners/anthropic/langchain_anthropic/chat_models.py (_make_message_chunk_from_anthropic_event):

1. The content_block_start handler added in 1.4.8 only emits the initial thinking content when thinking or signature is truthy. For an empty adaptive-thinking block, content_block_start carries thinking="" and signature="", so nothing is emitted and the thinking key is lost.

else:  # thinking
    thinking = getattr(event.content_block, "thinking", "") or ""
    signature = getattr(event.content_block, "signature", "") or ""
    if thinking or signature:   # <- empty block: start event is dropped entirely
        ...

2. The block is then reconstructed from deltas alone. A signature_delta produces {'signature': ..., 'type': 'thinking'} without a thinking key:

elif event.delta.type in {"thinking_delta", "signature_delta"}:
    content_block = event.delta.model_dump()
    content_block["index"] = event.index
    content_block["type"] = "thinking"
    message_chunk = AIMessageChunk(content=[content_block])

3. On replay, _format_messages passes the block through unchanged (it only filters keys, it does not restore thinking), so the invalid block reaches the API.

Possible fixes: always emit the content_block_start content for thinking blocks (including thinking=""), or default thinking to "" when building the block from a signature_delta. Restoring thinking: "" reconstructs the canonical form the API returned, so the signature still validates.

System Info

System Information

OS: Darwin
OS Version: Darwin Kernel Version 24.3.0: Thu Jan 2 20:24:16 PST 2025; root:xnu-11215.81.4~3/RELEASE_ARM64_T6000
Python Version: 3.12.10 (main, May 28 2025, 09:42:05) [Clang 17.0.0 (clang-1700.0.13.3)]

Package Information

langchain_core: 1.4.7
langchain: 1.3.9
langsmith: 0.8.15
deepagents: 0.6.10
langchain_anthropic: 1.4.8
langchain_aws: 1.4.1
langchain_google_genai: 4.2.5
langchain_mcp_adapters: 0.2.2
langchain_modal: 0.0.2
langchain_openai: 1.1.12
langchain_protocol: 0.0.15
langgraph_sdk: 0.4.2

Other Dependencies

anthropic: 0.109.1