Vercel AI SDK 6.0

Apply the Node.js 24 and TypeScript 7 prompts first. These rules cover Vercel AI SDK 6.0 provider configuration, prompts, tools and approvals, agent loops, structured output, streaming, persistence, MCP, and observability.

Providers, models, and call budgets

• Build model IDs, provider instances, API keys, base URLs, and request headers from server configuration rather than from a request body or a useChat transport option.
• Keep provider and gateway credentials out of client bundles, DefaultChatTransport headers, and any body or headers callback that runs in the browser.
• Allowlist the model IDs, modalities, and providerOptions a caller may select instead of spreading a caller-supplied settings object into generateText or streamText.
• Set maxOutputTokens on every call; 6.0 sends no output ceiling of its own when the setting is omitted.
• Bound each call with timeout using totalMs and stepMs alongside abortSignal, because maxRetries defaults to 2 and one request can reach the provider three times.
• Treat an abort or a browser disconnect as local cancellation only, and reconcile provider-side spend and partially completed side effects separately.

Prompts, messages, and validation

• Keep system and agent instructions text in server code, never let a UIMessage part supply, append to, or replace it, and set allowSystemInMessages: false on streamText, generateText, and ToolLoopAgent: validateUIMessages accepts a client-sent role: 'system' message, convertToModelMessages turns it into a real system message, and with