ai-sdk-core-v7

Skill: ai-sdk-core

Scope

Folder major tracks the installed ai package in package.json, not a skill-file revision
Node.js 22+, ESM; Zod 4 for inputSchema
Never write AI SDK from memory. Grep node_modules/ai/docs/ and node_modules/ai/src/ for the installed version. Else https://ai-sdk.dev/docs (append .md) or https://ai-sdk.dev/api/search-docs?q=
v6 leftovers: https://ai-sdk.dev/docs/migration-guides/migration-guide-7-0

generateText / streamText for request handlers and one-shot API routes; instructions for system-style prompt
Request-handler tool loops: stopWhen: isStepCount(n) on generateText / streamText — not maxSteps, not stepCountIs
ToolLoopAgent + generate() / stream() for durable agents — not hand-rolled loops, not new Agent()
UI streams: createUIMessageStreamResponse + toUIMessageStream({ stream: result.stream }) from 'ai'. Use these for new handlers and for dedicated stream/SDK updates
result.toUIMessageStreamResponse() still works in v7 (deprecated). Leave it only when the task is unrelated to streaming or the SDK
Tool results must be JSON-serializable (no Date)
System prompt lives in instructions. Client messages must not include { role: 'system' } unless the server fully trusts the payload

Read bundled docs before writing calls
convertToModelMessages when the client sends UI messages
Reject untrusted { role: 'system' } in messages. Do not set allowSystemInMessages: true for client-supplied chat

toDataStreamResponse / pipeDataStreamToResponse
stepCountIs / maxSteps / result.fullStream / top-level system: (use isStepCount, result.stream, instructions)
Guessing model IDs