SummarizationMiddleware fails - list object has no attribute strip

[tsmith-perchgroup]

Checked other resources

• 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-cli
• 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-perplexity
• langchain-prompty
• langchain-qdrant
• langchain-xai
• Other / not sure / general

Example Code (Python)

from langchain_openai import AzureChatOpenAI
import warnings 
from langchain.agents import create_agent
from langchain.agents.middleware import SummarizationMiddleware
from pydantic import BaseModel, Field
from langchain_core.messages import HumanMessage
from langchain_core.prompts import PromptTemplate
import json

warnings.filterwarnings('ignore')

azure_open_ai_config_gpt5 = {
    "api_key": "**obfuscated**",
    "azure_endpoint": "**obfuscated**",
    "azure_deployment": "gpt-5",
    "model": "gpt-5",
    "deployment_name": "gpt-5",
    "api_version": "2025-03-01-preview"
}


reasoning = ["minimal", 'low', 'medium', 'high']
verbosity = ['low', 'medium', 'high']
      
model = {
    (
        f"gpt5_{r}_reasoning" if v == "medium"
        else f"gpt5_{r}_reasoning_{v}_verbosity"
    ): AzureChatOpenAI(
        **azure_open_ai_config_gpt5,
        temperature=0,
        model_kwargs={"reasoning": {"effort": r}, "verbosity": v, "max_output_tokens": 128_000},
        timeout=60*10,
        max_retries=3,
        max_tokens=128_000
    )
    for r in reasoning for v in verbosity
}

def add(
  x: int,
  y: int
) -> int:
  """Add two integers"""
  return x + y

class AgentResponse(BaseModel):
  content: str = Field(..., description="Standard response")
  tool_calls: int = Field(..., description="Number of tool calls")

agent = create_agent(
  model=model['gpt5_medium_reasoning'],
  response_format=AgentResponse,
  middleware=[SummarizationMiddleware(
    model=model['gpt5_low_reasoning'],
    max_tokens_before_summary=2000,
    messages_to_keep=10,
    summary_prompt="Summarize the following tool-call history: {messages}"
  )
  ],
  tools=[add],
  system_prompt="You are a helpful assistant that can add numbers",
)

test = agent.invoke(
  {
    "messages": [
      HumanMessage(
        """
        Find the first 100 numbers of the fibonacci sequence using your tools
        """
      )
    ]
  }
)

Error Message and Stack Trace (if applicable)

'list' object has no attribute 'strip'"

Description

The error message surfaces in the returned message from the summarisation middleware:

HumanMessage(content="Here is a summary of the conversation to date:\n\nError generating summary: 'list' object has no attribute 'strip'", additional_kwargs={}, response_metadata={}, id='cce7000c-af83-4b57-8f3a-1d3e901e1720'),
  AIMessage(content=[{'arguments': '{"x":28657,"y":46368}', 'call_id': 'call_ChyA0TPDeU3LMnB14g2Prg1R', 'name': 'add', 'type': 'function_call', 'id': 'fc_081ee22ba7dbf85500691f33b3ea0881939f076de38f38f90c', 'status': 'completed'}], additional_kwargs={}, response_metadata={'id': 'resp_081ee22ba7dbf85500691f33b356f081938b4355be25524dcd', 'created_at': 1763652531.0, 'metadata': {}, 'model': 'gpt-5', 'object': 'response', 'service_tier': 'auto', 'status': 'completed', 'model_provider': 'openai', 'model_name': 'gpt-5'}, id='resp_081ee22ba7dbf85500691f33b356f081938b4355be25524dcd', tool_calls=[{'name': 'add', 'args': {'x': 28657, 'y': 46368}, 'id': 'call_ChyA0TPDeU3LMnB14g2Prg1R', 'type': 'tool_call'}], usage_metadata={'input_tokens': 4418, 'output_tokens': 23, 'total_tokens': 4441, 'input_token_details': {'cache_read': 4352}, 'output_token_details': {'reasoning': 0}})

I can see from the mlflow trace that the summary is successfully being created:

User asked for the first 100 Fibonacci numbers.
The assistant iteratively used a single tool named "add" to sum consecutive terms, starting with add(0, 1) = 1.
Subsequent calls summed the previous two results: add(1, 1) = 2, add(1, 2) = 3, add(2, 3) = 5, … continuing up to add(17711, 28657) = 46368.
Results produced: 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610, 987, 1597, 2584, 4181, 6765, 10946, 17711, 28657, 46368.
The process generated 23 Fibonacci numbers and stopped at 46368, not completing the requested 100 numbers.

Clearly somewhere a list is attempting to be processed as a string but I was wondering if anyone else has a similar issue? I don't think I'm doing anything wrong here but if there's a fix/workaround that would be great appreciated.

System Info

System Information

OS: Linux
OS Version: #100-Ubuntu SMP Tue May 27 21:41:06 UTC 2025
Python Version: 3.12.3 (main, Aug 14 2025, 17:47:21) [GCC 13.3.0]

Package Information

langchain_core: 1.0.7
langchain: 1.0.8
langsmith: 0.4.44
langchain_openai: 1.0.3
langchain_tavily: 0.2.12
langchain_text_splitters: 1.0.0
langgraph_sdk: 0.2.9

Optional packages not installed

langserve

Other Dependencies

aiohttp: 3.13.2
httpx: 0.27.0
jsonpatch: 1.33
langgraph: 1.0.3
openai: 2.8.1
opentelemetry-api: 1.32.1
opentelemetry-sdk: 1.32.1
orjson: 3.11.4
packaging: 24.2
pydantic: 2.10.6
pytest: 8.3.5
pyyaml: 6.0.2
requests: 2.32.3
requests-toolbelt: 1.0.0
rich: 14.1.0
tenacity: 9.0.0
tiktoken: 0.12.0
typing-extensions: 4.12.2
zstandard: 0.25.0

[Manju080]

/assign

I’ve gone through the problem description and I’m comfortable implementing the changes.

[tsmith-perchgroup]

@Manju080 Is there an interim workaround?

[Manju080]

@Manju080 Is there an interim workaround?

Yes, there is a workaround for now.
The failure happens because SummarizationMiddleware eventually formats the collected messages into the summary prompt, but the tool-call history contains AIMessage.content values that are lists instead of strings, and the template logic calls .strip() on them. That’s where the 'list' object has no attribute 'strip' error comes from.

[keenborder786]

@tsmith-perchgroup You could use a custom CustomSummarizationMiddleWare:

from langchain.agents.middleware import SummarizationMiddleware
from langchain_core.messages import HumanMessage
from langchain_core.messages import AnyMessage

class CustomSummarizationMiddleWare(SummarizationMiddleware):
    def _create_summary(self, messages_to_summarize: list[AnyMessage]) -> str:
        """Generate summary for the given messages."""
        if not messages_to_summarize:
            return "No previous conversation history."

        trimmed_messages = self._trim_messages_for_summary(messages_to_summarize)
        if not trimmed_messages:
            return "Previous conversation was too long to summarize."

        try:
            response = self.model.invoke(self.summary_prompt.format(messages=trimmed_messages))
            if isinstance(response, list):
                response = response[0]
            return response.text.strip()
        except Exception as e:  # noqa: BLE001
            return f"Error generating summary: {e!s}"

But will have to do a RCA as to why .model.invoke returns a list in your case.

[keenborder786]

If possible can you add a print response immediately after response and show me the what print is coming to. .invoke should always return a instance AIMessage so it's strange as to why list is coming in your case.

[tsmith-perchgroup]

Great that's working thanks.

Sidenotes:

• Presumably the SummarizerMiddleware preserves tool call <-> toll message pairs? I've had problems with trimming message histories in the past due to severed pairs.
• I'm also finding that setting my max_tokens_before_summary > 2000 just gives me:

Previous conversation was too long to summarize.

Why is this? The entire tool chain without any summarisation from the above is only 50,000 tokens, which these models can handle easily. Obviously _trim_messages_for_summary is returning None or equivalent, but why is this?

print(response) example:

content=[{'id': 'rs_09334401da20943900691f49fbe68c81978b227c59a09f2b6f', 'summary': [], 'type': 'reasoning'}, {'type': 'text', 'text': '- Goal: Generate the first 100 Fibonacci numbers via iterative tool calls.\n- Method: Repeatedly used an add(x, y) tool to sum the last two Fibonacci numbers, starting from 0 and 1.\n- Results produced (23 tool calls): 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610, 987, 1597, 2584, 4181, 6765, 10946, 17711, 28657, 46368.\n- Progress: Reached Fibonacci number F24 (if F0=0, F1=1). Still short of the requested 100 numbers.', 'annotations': [], 'id': 'msg_09334401da20943900691f4a020b388197be3400443bf132d0'}] additional_kwargs={} response_metadata={'id': 'resp_09334401da20943900691f49fb49f081979bef44d0495b6d7a', 'created_at': 1763658235.0, 'metadata': {}, 'model': 'gpt-5', 'object': 'response', 'service_tier': 'default', 'status': 'completed', 'model_provider': 'openai', 'model_name': 'gpt-5'} id='resp_09334401da20943900691f49fb49f081979bef44d0495b6d7a' usage_metadata={'input_tokens': 9112, 'output_tokens': 934, 'total_tokens': 10046, 'input_token_details': {'cache_read': 0}, 'output_token_details': {'reasoning': 768}}

[keenborder786]

okay thank you @tsmith-perchgroup, will need to do a better RCA.

• Yes that is indeed a problem since sometimes _determine_cutoff_index fail to find a safe cutoff for AI and Tool Call Pair
• For this can you set trim_tokens_to_summarize to None

[tsmith-perchgroup]

Last questions -

• what function does trim_tokens_to_summarize parameterise?
• is there any way to inject the message as a SystemMessage, rather than as a HumanMessage?