Vercel AI SDK 7 shipped breaking changes in patch releases
A renamed option and a new finish reason can break caret-pinned builds on update. Plus Harnesses now wrap Claude Code and Codex, and Kimi K3.1's weights near.
Copy markdown
Patch bumps that break the build
`ai@7.0.144` renamed `experimental_lowMemory` to `experimental_streamTextLowMemory()` — the old name no longer compiles on 7.0.154 — and 7.0.149 added an `interrupted` FinishReason that trips exhaustive switches. Both shipped as patch releases, so a `^7.0` range pulls them in on `npm update`. Pin an exact version and audit before bumping.
Harnesses now wrap real coding agents
`HarnessAgent` runs Claude Code, Codex, Cursor, GitHub Copilot, Grok Build and Pi behind the same `generate()`/`stream()` as a model call — the docs now list around ten adapters, up from three at June's canary. Swap `claudeCode` for `codex` without touching your `useChat` UI. Caveats: no adapter exposes session history yet, and structured output is unsupported on Cursor, Copilot and Pi.
Tool Search trims the first prompt
`toolSearch()` with `deferLoading: true` keeps big tool catalogs out of the opening prompt: the model sees one search tool, and the top matches become callable on the next step (custom ranking since 7.0.127, MCP supported). Catch: set `stopWhen` explicitly, since `generateText`/`streamText` default to one step and the matched tools would never fire.
Resumable streams and per-step saves
`streamRetries` reruns only the dropped model step instead of the whole generation, and mid-stream provider errors now normalize to a single `StreamProviderError`. `onUIMessageStepEnd` (7.0.132) persists chat state after each agent step so a crash mid-run doesn't wipe it, and `WebSocketChatTransport` (7.0.135) moves `useChat` onto one persistent socket.
Kimi K3.1 Pro: frontier-class open weights land Nov 4
Moonshot's model is live on its API at $2/$10 per million tokens — about half Claude Opus 5.5 — and scored 55 on Artificial Analysis, three points off Opus 5.5 (58) and ahead of GPT-6 Astra and Gemini 4 Argon (53). Full open weights are promised Nov 4, so self-hosters can start sizing capacity now.