Who contributed to Vercel on July 17, 2026?

13 developers shipped this update, including Lars Grammel, Nick Oates, Walter Korman, Greg Schofield, Bryan Hunter, Fantix King, msullivan, and NathanColosimo, and 5 more.

What were the notable Vercel updates?

docs: document AI SDK 6 removal of `name` from function tool definitions (#17316), feat: pass end-user identifiers to xAI video generation and editing requests, and feat: add kimi-k3 model and `reasoningEffort` provider option.

AI SDK 6 BREAKING CHANGE SURFACES WHILE TURBOPACK INLINES RUNTIME BOOTSTRAPPING

AI SDK 6's removal of the `name` parameter from function tool definitions is now documented after breaking upgraders, while Next.js Turbopack merges runtime optimization PRs that eliminate per-route overhead.

The AI SDK team shipped breaking documentation [1] clarifying that AI SDK 6 no longer accepts a `name` field in function tool definitions, surfacing type errors on upgrade. This lands after developers hit unclear TypeScript overload errors when migrating from SDK 5. Separately, the same desk added support for xAI video generation user identifiers [2], Kimi K3 reasoning models [3], and next.config.ts migrations [4] across examples. On the platform side, Vercel CLI gained explicit `--project` flags across environment, firewall, and route commands [5], letting teams target projects without directory linking. Python functions now always run `compileall` instead of heuristic checks [6], simplifying bytecode packing and preventing accidental size limit violations. Over in Next.js, a stack of Turbopack commits consolidates runtime code into a shared asset [7], adds `registerEntry()` for bootstrap handling [8], inserts chunk parameters into the build manifest [9], and finally inlines the bootstrap inline to drop per-route runtime files entirely [10], which together eliminate payload duplication across routes. Workflow fixed dotted tsconfig alias discovery that was skipping workflow imports [11], reorganized world documentation under a versioned `/worlds` canonical home [12], and added optional `getMany()` batch reads to avoid N+1 storage queries [13]. Model evals confirmed Kimi K3 hits 92 percent base (96 with docs) [14] and cleaned up bogus GPT 5.6 runs that were scoring against fallback metadata [15].