import { createChat } from "@shadcn/helpers/ai-sdk"

const chat = createChat()
  .user("What's the weather in San Francisco?")
  .sleep(800)
  .assistant(({ writer }) => {
    writer.reasoning("I should call the weather tool first.")
    writer
      .tool("getWeather", { input: { city: "San Francisco" } })
      .sleep(1200)
      .output({ temperature: 72, condition: "sunny" })
    writer.text("It's sunny and 72 degrees in San Francisco.")
  })

Human in the loop

A tool call left unresolved pauses the turn: the input streams, the turn finishes, and the client decides what happens next. A callback turn scripted after a paused turn becomes a continuation. It materializes when the follow-up request arrives and receives the resolved toolCall in its context.

For client-executed tools, the user supplies the output through addToolOutput and the continuation reads it:

const chat = createChat()
  .user("Help me plan the release.")
  .assistant(({ writer }) => {
    writer.tool("askQuestions", { dynamic: true, input: { questions } })
  })
  .assistant(({ writer, toolCall }) => {
    writer.text(`Got it. Starting with ${toolCall?.output?.answers.direction}.`)
  })

For approval-gated tools, needsApproval pauses behind the user's decision. output (or errorText) then means "stream this if approved"; denial streams tool-output-denied automatically:

const chat = createChat()
  .user("Clean up old deployments.")
  .assistant(({ writer }) => {
    writer.text("This will delete 3 deployments. Approve?")
    writer.tool("deleteDeployments", {
      input: { count: 3 },
      needsApproval: true,
      output: { deleted: 3 },
    })
  })
  .assistant(({ writer, toolCall }) => {
    writer.text(
      toolCall?.approved
        ? "Deleted 3 deployments."
        : "Okay, leaving them in place."
    )
  })

Wire the client with the AI SDK's own triggers: sendAutomaticallyWhen: lastAssistantMessageIsCompleteWithToolCalls for client-executed tools, or lastAssistantMessageIsCompleteWithApprovalResponses plus addToolApprovalResponse({ id: part.approval.id, approved }) for approvals.

Continuation callbacks must be pure; regenerating re-resolves them against the current transcript. get() stops before the first continuation turn, since a continuation has no message without a live transcript.

Installation

npm install @shadcn/helpers

Documentation

Visit https://ui.shadcn.com/docs to view the documentation.

Contributing

Please read the contributing guide.

License

Licensed under the MIT license.