The Wire · Showcase
TURBOPACK CUTS BUILD CACHE MISSES, AI SDK GRADUATES EXPERIMENTAL FEATURES
By RepoJournal · Filed · About Vercel
Next.js shipped two critical Turbopack optimizations that slash task cache misses by 74%, while Vercel AI SDK promoted five experimental APIs to stable—the biggest API surface shift since launch.
The Turbopack team is methodically cutting task overhead. First, they eliminated `ReferencedAsset::from_resolve_result` as a turbotask [1], a change that addresses a task that was executing 1.9M times with only a 74% hit rate. Then they simplified asset ident constructors [2], converting per-method turbotasks into plain Rust builders with a single final `into_vc()` call—less ceremony, same output, faster caches. These changes compound: fewer turbotasks mean fewer scheduler decisions, faster builds, and cleaner compilation graphs. Request store handling got a cleaner API too [3], decoupling store creation from HTTP primitives so consumers can build stores from serializable inputs instead of concrete `req`/`res` pairs. Meanwhile, Vercel AI SDK is graduating experimental features: `include` is now stable [7], replacing `experimental_include` for controlling response content; tool metadata has a dedicated path [4] instead of merging into provider metadata; and deprecated callbacks [6] are back for SDK 6 compatibility. The `includeRawChunks` option moved to `include.rawChunks` [5]. On Turborepo, two security hardening PRs landed: local devtools WebSocket now requires per-session token validation [8] instead of permissive CORS, and OTel endpoint validation [9] rejects malformed URLs before initializing exporters. CLI got two small wins [10] skips SAML-limited teams during `vercel link` auto-detection, and [11] adds `--error` flag for manual deployment continuation.
Action items
- → Rebuild Next.js projects with latest main to benefit from Turbopack cache improvements vercel/next.js [plan]
- → Update Vercel AI SDK code: migrate experimental_include → include, remove deprecated experimental_onToolCall* if upgrading beyond SDK 6 vercel/ai [plan]
- → Verify local Turborepo devtools work with new token-based WebSocket auth if you're on main vercel/turborepo [monitor]
References
- [1] Drop `ReferencedAsset::from_resolve_result` as a turbotask (#93297) vercel/next.js
- [2] Turbopack: simplify asset ident constructors (#93213) vercel/next.js
- [3] Refactor: Decouple request store creation from `req` / `res` ↗ vercel/next.js
- [4] feat(ai): add toolMetadata for tool specific metdata (#15021) vercel/ai
- [5] feat: move includeRawChunks to include.rawChunks (#15059) vercel/ai
- [6] feat: add deprecated tool call lifecycle callback aliases for AI SDK 6 compatibility (#15062) vercel/ai
- [7] feat: make include stable (#15063) vercel/ai
- [8] fix: Authenticate local devtools WebSocket ↗ vercel/turborepo
- [9] fix: Reject OTel endpoints with userinfo ↗ vercel/turborepo
- [10] Skip limited teams during vercel link auto detect (#16220) vercel/vercel
- [11] [cli] Allow continuing a deployment to error ↗ vercel/vercel
FAQ
- What changed in Vercel on May 7, 2026?
- Next.js shipped two critical Turbopack optimizations that slash task cache misses by 74%, while Vercel AI SDK promoted five experimental APIs to stable—the biggest API surface shift since launch.
- What should Vercel teams do about it?
- Rebuild Next.js projects with latest main to benefit from Turbopack cache improvements • Update Vercel AI SDK code: migrate experimental_include → include, remove deprecated experimental_onToolCall* if upgrading beyond SDK 6 • Verify local Turborepo devtools work with new token-based WebSocket auth if you're on main
- Which Vercel repositories shipped on May 7, 2026?
- vercel/next.js, vercel/ai, vercel/turborepo, vercel/vercel