Lesson 4: BookingStep — Pick a doctor, book, end

Prompt with carried state

BookingStep pulls state from SymptomsStep before prompting:

public getPrompt(): string {
  const doctors = this.getState('doctors');
  const symptoms = this.getState('symptoms');
  return `You have the user's symptoms: ${symptoms}. Here are the available doctors: ${JSON.stringify(doctors)}. Help the user pick a doctor and an available time slot. Once they decide, use the 'book_appointment' tool.`;
}

State isolation means only the booking memory is loaded here—no leakage from earlier chat turns.

Tool + handler

public defineTool(): ToolType[] {
  return [
    {
      name: 'book_appointment',
      description: 'Book an appointment with the selected doctor at the chosen time.',
      schema: z.object({
        doctorName: z.string().describe('Name of the selected doctor'),
        timeSlot: z.string().describe('The chosen time slot'),
      }),
    },
  ];
}

protected async book_appointment(tool: ToolCall): Promise<ToolResponseType> {
  const { doctorName, timeSlot } = tool.args;
  this.saveState({ doctorName, timeSlot, booked: true });
  return { step: EndStep, tool: `Successfully booked with ${doctorName} at ${timeSlot}. Thank you!` };
}

Another narrow tool surface: only booking + end_chat.
Booking persists confirmation fields and jumps directly to EndStep.

End the flow

src/myflow/medical-flow/medical-flow.ts wires the sequence:

return [
  new SymptomsStep(this, true).setTemperature(0.5).useModel(model),
  new BookingStep(this, false).useModel(model),
  new EndStep(this).useMemory('end').useModel(model),
];

Explicit order keeps debugging simple (one list, deterministic transitions).
EndStep uses its own memory slice (end) so the goodbye prompt stays clean.
Temperature is relaxed on symptoms intake, then the rest reuses the flow default.

[!Pattern] This is the “tool-controlled step switch” pattern: tools decide when to exit the current step and which step to activate next, giving you traceable, testable control flow.