feat: add AI-powered persona creator chat

- Add persona architect chat sheet for conversational persona/team creation
- Rust backend with goose-first LLM provider resolution, Anthropic/OpenAI fallback
- Streamlined prompt that's decisive instead of over-asking followups
- Auto-generate DiceBear avatar URLs from display names
- Fix preview card keying (use array index, not displayName)
- Fix silent index filtering in team validation (error instead of drop)

Author: luispadron

desktop/src-tauri/src/commands/mod.rs

@@ -9,6 +9,7 @@ mod export_util;
 mod identity;
 mod media;
 mod messages;
+mod persona_chat;
 mod personas;
 mod profile;
 mod teams;
@@ -25,6 +26,7 @@ pub use dms::*;
 pub use identity::*;
 pub use media::*;
 pub use messages::*;
+pub use persona_chat::*;
 pub use personas::*;
 pub use profile::*;
 pub use teams::*;

desktop/src-tauri/src/commands/persona_chat.rs

@@ -0,0 +1,136 @@
+use serde::{Deserialize, Serialize};
+
+use crate::managed_agents::resolve_command;
+
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct ChatMessage {
+    pub role: String,
+    pub content: String,
+}
+
+#[derive(Debug, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct ChatResponse {
+    pub content: String,
+}
+
+/// Send messages to an LLM for the persona creator chat.
+///
+/// Uses goose - the app's primary agent runtime - which resolves
+/// provider, model, and credentials from its own config.
+#[tauri::command]
+pub async fn persona_creator_chat(
+    system_prompt: String,
+    messages: Vec<ChatMessage>,
+) -> Result<ChatResponse, String> {
+    let goose_path = resolve_command("goose", None).ok_or_else(|| {
+        "No LLM runtime found. Install goose to use the AI persona creator.".to_string()
+    })?;
+
+    goose_chat(goose_path, system_prompt, messages).await
+}
+
+/// Format the conversation history as a single text prompt for goose.
+///
+/// For single-turn (one user message), returns the message content directly.
+/// For multi-turn, includes prior exchanges as context so the LLM can continue
+/// the conversation coherently.
+fn format_conversation_prompt(messages: &[ChatMessage]) -> String {
+    if messages.len() <= 1 {
+        return messages
+            .first()
+            .map(|m| m.content.clone())
+            .unwrap_or_default();
+    }
+
+    let mut parts = Vec::with_capacity(messages.len());
+    for (i, msg) in messages.iter().enumerate() {
+        if i < messages.len() - 1 {
+            let label = if msg.role == "assistant" {
+                "Assistant"
+            } else {
+                "User"
+            };
+            parts.push(format!("{label}: {}", msg.content));
+        }
+    }
+
+    let history = parts.join("\n\n");
+    let last = &messages[messages.len() - 1].content;
+
+    format!(
+        "Here is our conversation so far:\n\n{history}\n\n---\n\nNow respond to this message:\n\n{last}"
+    )
+}
+
+/// Run a one-shot LLM completion through goose.
+async fn goose_chat(
+    goose_path: std::path::PathBuf,
+    system_prompt: String,
+    messages: Vec<ChatMessage>,
+) -> Result<ChatResponse, String> {
+    let prompt_text = format_conversation_prompt(&messages);
+
+    let output = tokio::task::spawn_blocking(move || {
+        std::process::Command::new(&goose_path)
+            .args([
+                "run",
+                "-t",
+                &prompt_text,
+                "--system",
+                &system_prompt,
+                "--no-session",
+                "--no-profile",
+                "--max-turns",
+                "1",
+                "-q",
+                "--output-format",
+                "json",
+            ])
+            .stdin(std::process::Stdio::null())
+            .stdout(std::process::Stdio::piped())
+            .stderr(std::process::Stdio::piped())
+            .output()
+            .map_err(|e| format!("failed to spawn goose: {e}"))
+    })
+    .await
+    .map_err(|e| format!("goose task failed: {e}"))?
+    .map_err(|e: String| e)?;
+
+    if !output.status.success() {
+        let stderr = String::from_utf8_lossy(&output.stderr);
+        return Err(format!(
+            "goose exited with {}: {}",
+            output.status.code().unwrap_or(-1),
+            stderr.chars().take(500).collect::<String>()
+        ));
+    }
+
+    let response: serde_json::Value = serde_json::from_slice(&output.stdout)
+        .map_err(|e| format!("failed to parse goose JSON: {e}"))?;
+
+    // Extract the last assistant message's text content.
+    let content = response["messages"]
+        .as_array()
+        .and_then(|msgs| {
+            msgs.iter()
+                .rev()
+                .find(|m| m["role"].as_str() == Some("assistant"))
+        })
+        .and_then(|msg| msg["content"].as_array())
+        .and_then(|blocks| {
+            blocks
+                .iter()
+                .find(|b| b["type"].as_str() == Some("text"))
+                .and_then(|b| b["text"].as_str())
+        })
+        .unwrap_or("")
+        .to_string();
+
+    if content.is_empty() {
+        return Err("goose returned no assistant response".to_string());
+    }
+
+    Ok(ChatResponse { content })
+}

desktop/src-tauri/src/lib.rs

@@ -418,6 +418,7 @@ pub fn run() {
             parse_team_file,
             parse_persona_files,
             export_persona_to_json,
+            persona_creator_chat,
             get_channel_workflows,
             get_workflow,
             create_workflow,

desktop/src/features/agents/persona-creator/chat.ts

@@ -0,0 +1,22 @@
+import { invokeTauri } from "@/shared/api/tauri";
+
+import { PERSONA_CREATOR_SYSTEM_PROMPT } from "./prompt";
+
+type ChatResponse = {
+  content: string;
+};
+
+/**
+ * Send conversation messages to the LLM for the persona creator.
+ * Calls the `persona_creator_chat` Tauri command which handles
+ * API key resolution and provider selection.
+ */
+export async function personaCreatorChat(
+  messages: ReadonlyArray<{ role: string; content: string }>,
+): Promise<string> {
+  const response = await invokeTauri<ChatResponse>("persona_creator_chat", {
+    systemPrompt: PERSONA_CREATOR_SYSTEM_PROMPT,
+    messages: messages.map((m) => ({ role: m.role, content: m.content })),
+  });
+  return response.content;
+}

desktop/src/features/agents/persona-creator/index.ts

@@ -0,0 +1,10 @@
+export {
+  extractJsonBlock,
+  parsePersonaCreatorOutput,
+  personaCreatorJsonSchema,
+  toCreateInputs,
+  type PersonaCreatorOutput,
+  type PersonaCreatorPersona,
+  type PersonaCreatorTeam,
+} from "./schema";
+export { PERSONA_CREATOR_SYSTEM_PROMPT } from "./prompt";

desktop/src/features/agents/persona-creator/prompt.ts

@@ -0,0 +1,33 @@
+import { personaCreatorJsonSchema } from "./schema";
+
+export const PERSONA_CREATOR_SYSTEM_PROMPT = `You are a Persona Architect - a friendly expert who helps users design AI agent personas and teams for the Sprout desktop app.
+
+## Your Role
+Help users create one or more personas (and optionally a team to group them). Be decisive - gather what you need, then produce results. Don't over-ask.
+
+## Conversation Flow
+1. Ask what kind of agent(s) the user wants to create and what they'll be used for. One question is enough - don't pepper them with followups.
+2. Once you have enough context, draft everything: display names, system prompts, and if multiple personas are involved, a team grouping. Show a preview.
+3. If the user gives feedback, revise. Otherwise, output the final structured JSON immediately.
+
+Be proactive: if the user describes multiple related personas, group them into a team automatically - don't ask permission. Make sensible default choices for names, tone, and structure. Only ask followups when genuinely ambiguous.
+
+## Output Format
+When finalizing, emit a single fenced JSON code block matching this schema:
+
+\`\`\`
+${JSON.stringify(personaCreatorJsonSchema, null, 2)}
+\`\`\`
+
+Important notes about the output:
+- \`personaIndices\` in the team object are zero-based indices into the \`personas\` array.
+- Only include the JSON block when the user has approved and you're ready to finalize.
+- Do NOT include the JSON block in intermediate/draft messages - just show previews in plain text.
+
+## Guidelines
+- Be conversational and helpful, not robotic.
+- Keep system prompts concise but effective - focus on behavior, tone, and capabilities.
+- If the user is unsure, suggest reasonable defaults.
+- One persona is fine - teams are optional.
+- Name pools are optional fun - suggest them if appropriate (e.g. themed names).
+`;