[MikeAlhayek]

What

A new, opt-in AI persona feature describes reusable characters: who the assistant is, the name it uses, its language, its avatar and the voice it speaks with. Greetings stay with the profile's welcome message, so one persona can serve profiles that open differently. An AI profile, a chat interaction or a profile template links to one optional persona by its technical name (AIPersonaMetadata.PersonaId). The persona is linked, not copied, so editing it changes everything that uses it from the next message.

A persona answers who; the system message and prompt templates still answer what.

Its own catalog, not an AI template

A template is a blueprint that is copied when it is used; a persona stays linked. So personas have their own catalog instead of being AI templates with a Persona source:

• AIPersona is a named catalog item (instructions, assistant name, language, voice and voice deployment, avatar URL). IAIPersonaManager merges saved personas with every IAIPersonaProvider: persona files in the content root's Templates/Personas, or embedded ones registered with AddEmbeddedPersonas(assembly). A saved persona replaces a file with the same name. The catalog handler requires a unique name and instructions, and an update never changes the name.
• Opt-in: .AddPersonas(personas => personas.AddYesSqlStores()) (or .AddEntityCoreStores()) on the AI suite builder. YesSql hosts create the index table with CreateAIPersonaIndexSchemaAsync. Without the feature, persona links are ignored and chats run with the profile's own instructions and voice.

How it works

• IAIPersonaResolver is the single place that turns a link into a ResolvedAIPersona (DefaultAIPersonaResolver). A missing persona is logged and the request carries on without one. A Liquid error drops the instructions but keeps the rest.
• Text on every request. AIPersonaCompletionContextBuilderHandler runs once the profile or interaction has built its system message and places the persona block first, then the prompt templates, then the system message, so business rules win a conflict. When the persona has an assistant name, the block starts with "Your name is …", and the instructions can use it as the Liquid variable assistant_name. The persona is also exposed under AICompletionContextKeys.Persona. Realtime sessions read their instructions from the same completion context.
• Voice when a session opens. SelectVoiceAsync chooses the explicit voice (caller, profile or interaction), then the persona's voice, then the deployment default. The persona's voice is used only when the speaking deployment has the same provider as the deployment it was picked from. The chat hubs go through IServiceProvider.ResolvePersonaAsync / SelectPersonaVoiceAsync, so a host without personas keeps working.
• Name and avatar in chat. @crestapps/ai-chat-ui takes a new assistantAvatarUrl setting (data-coreai-chat-assistant-avatar-url) beside assistantLabel. Assistant messages and the typing indicator show the persona's name and a round avatar in place of the robot icon; without them the chat looks as before. Hosts get the values from IServiceProvider.ResolveChatAssistantDisplayAsync(resource), which returns the default look when personas are not enabled.
• Saving. PersonaChatInteractionSettingsHandler reads personaId; a client that omits the key keeps the current link. Profile JSON accepts a top-level PersonaId. AIChatSession.PersonaId records the persona a session started with.

Sample hosts

Both the MVC (YesSql) and Blazor (EF Core) hosts add personas:

• a Personas admin page and dashboard card to list, create, edit and delete personas (a file-based persona opens as a copy that replaces the file once saved);
• an AI persona picker on AI profiles and Profile templates, hidden when personas are not enabled;
• the persona's name and avatar on assistant messages in the chat page and the widget.

Both hosts were checked live: personas created through each store, linked to profiles, and chats answered in character with the persona's name and avatar.

New public types

AIPersona, ResolvedAIPersona, IAIPersonaManager / DefaultAIPersonaManager, IAIPersonaProvider with AIPersonaFileSystemProvider and EmbeddedResourceAIPersonaProvider, AIPersonaFileParser, IAIPersonaResolver / DefaultAIPersonaResolver, AIPersonaMetadata, AIPersonaExtensions, AIPersonaServiceProviderExtensions, AIPersonaResolverExtensions, AIChatAssistantDisplay, CrestAppsAIPersonasBuilder, PersonaChatInteractionSettingsHandler, AICompletionContextKeys.Persona, AIChatSession.PersonaId, and the YesSql AIPersonaIndex.

Tests and docs

• Tests cover the resolver, the persona handler, the manager's merge rules, catalog validation, file parsing and providers, the opt-in registration, the chat display mapping, and both sample view models.
• Full suite passes locally (4,722). The solution builds in Release with -warnaserror. npm run lint, npm run check and test:unit pass, and the rebuilt wwwroot assets are committed.
• Docs: AI Personas page, an Assistant name and avatar section on the Chat page, and 2.0.0 changelog entries.

The Orchard Core side (persona editor, pickers, built-in persona files, recipe step, chat views) follows on a CrestApps.OrchardCore branch once this ships in a preview.

🤖 Generated with Claude Code