Pouchy — Leo

Persona — template defaults applied to every end user's instance

Name

Persona archetype

System prompt

Initial relationship stage

Model tier

Image generation · the agent can generate real images on request (generate_image tool; the result is a link in the reply)

Content safety

Custom refusal message (optional)

Content safety runs an OpenAI moderation check on each end-user message in parallel with the turn (no added latency; fails open on outages). Standard blocks only the categories no deployment should serve (minors, self-harm instructions, threats…) and leaves a romance companion's normal range untouched; Strict additionally blocks all sexual/hate/harassment/violence content for family-friendly apps. Blocked turns reply with the refusal line, are marked in the user transcript, and land in the audit log (never with message content).

Which LLM class serves this agent's chat turns — matches the pricing page's Standard + Pro model rates. Pro turns bill at the Pro credit multiplier; pick Standard to control cost. Takes effect as each end user's next session starts.

The initial relationship stage only applies to newly created end-user instances; existing users' relationship progress is never reset.

Public profile — DeJoy / info page (end-user-facing)

Shown to end users on the agent's info/profile page (e.g. DeJoy or a Telegram/Discord-style client) and on the public AgentCard. Discovery-only — it never changes runtime behavior or bumps the template version.

Proactivity — let your backend bring users back after they go quiet (webhook notification)

UI preview — sign in to enable: after N silent hours, an agent.nudge event is pushed to your webhook.

The platform can't reach the end user's device, but your backend can: the event carries external_user_id and the silence duration; you choose the re-engagement channel (push, email, ...) and can mint a session on the spot for the "welcome back" chat. At most one nudge per silent period; requires a project webhook subscribed to agent.nudge (or '*').

Cost / latency alerts

Evaluated hourly over this agent's run traces. Subscribe a project webhook to the agent.cost_alert event to receive breaches; while still breached it re-notifies once per window.