Skip to main content

Settings

export interface AgentSettings {
  details: DetailsSettings;
  security: SecuritySettings;
  generativeAi: GenerativeAiSettings;
  conversationStart: ConversationStartSettings;
}

Four members, one per tab of the Agent settings dialog. The structure deliberately mirrors the dialog rather than flattening it, so a capture of one tab maps onto exactly one member.

Walked in Phase 4.

Details​

export interface DetailsSettings {
  displayName: string;
  description?: string;
  /** Read-only after the first save. */
  schemaName: string;
  /** Read-only after the first save. */
  agentId: Guid;
}

Two of the four fields are rendered disabled. They are modelled anyway because the tab displays them

  • omitting them would misrepresent the surface. See Identity for why the duplication with AgentIdentity is intentional.

Security​

export type AuthenticationMode =
  | 'NoAuthentication'
  | 'AuthenticateWithMicrosoft'
  | 'Manual';

export interface SecuritySettings {
  authentication: AuthenticationMode;
  requireSecuredAccess: boolean;
}

Raybot uses AuthenticateWithMicrosoft with requireSecuredAccess: false.

Do not enable requireSecuredAccess while relying on the Demo or Preview surfaces. Doing so locks the maker out of their own agent preview: the preview surface is not an authenticated channel, so it cannot satisfy the requirement, and the agent stops responding there with no diagnostic pointing back at the setting. Leave it off until a real channel is carrying traffic.

This is recorded in the TSDoc on the field itself, so it reaches anyone reading the type or the emitted schema.

Generative AI​

export type ModerationLevel =
  | 'Minimum'
  | 'Low'
  | 'Medium'
  | 'High'
  | 'Maximum';

export const MODERATION_LEVELS: readonly ModerationLevel[] = [
  'Minimum', 'Low', 'Medium', 'High', 'Maximum',
] as const;

export interface GenerativeAiSettings {
  moderationLevel: ModerationLevel;
  userFeedbackEnabled: boolean;
}

MODERATION_LEVELS exists as an ordered array alongside the union because the order is semantic: the values run from least to most restrictive, and the control is a slider, not a dropdown. A Set or an unordered union loses the fact that High is more restrictive than Medium.

Anything that renders the slider, or compares two agents' moderation posture, needs the ordering:

import { MODERATION_LEVELS } from '@raybot/object-model';

const isStricter = (a: ModerationLevel, b: ModerationLevel) =>
  MODERATION_LEVELS.indexOf(a) > MODERATION_LEVELS.indexOf(b);

Raybot: Medium, the midpoint, with userFeedbackEnabled: false.

Feedback is off because thumbs-up and thumbs-down chips would have appeared in every Preview capture, adding UI that carries no information about the agent while cluttering the gallery.

Conversation start​

export interface ConversationStartSettings {
  greeting?: string;
  suggestedPrompts: string[];
}

Both empty for Raybot. greeting is undefined; suggestedPrompts is [].

The asymmetry is intentional and follows the same convention used throughout the model: suggestedPrompts is a required array because an empty list is a meaningful, observed state, while greeting is optional because the field can genuinely be unset.

Leaving both empty was a testing decision. A greeting occupies the first turn of every Preview conversation, and suggested prompts bias the tester toward questions the agent is known to handle. Both would have contaminated the Phase 5 grounding evidence. An empty conversation start means the first thing in every transcript is a real question.

What is not modelled​

The dialog has exactly four tabs and all four are modelled. What is not captured is the Dataverse-side representation: these settings persist to columns on the bot row, and the model does not record which column backs which field. That mapping was never established during the audit. See API ground truth for what the Dataverse surface does expose.