Instinct layer: gestures before the reply

Mynah can react physically before the brain finishes composing its answer. A fast decision model watches the live transcript and picks one of the avatar's baked gesture clips — nod, wave, shake, laugh, shy, wink — or nothing. The model that talks does not move; the model that moves does not talk. The split comes from open-annie and runs on Jev (TypeSafe System One), with a local rules provider as the floor and the fallback.

How it works

1. Clips. Every gesture is a short pre-baked clip in the same frame domain as the avatar's idle loop (see Avatars). Bake them from the console's bake API with kind: "action" and a LivePortrait driving template:

   curl -sk -X POST https://<host>:9443/api/v1/avatars/bake \
     -H "authorization: Bearer $TOKEN" -H "content-type: application/json" \
     -d '{"avatar_id":"luna","kind":"action","action":"shake","motion_ref":"builtin:shake_face","apply":true}'

   Built-in templates that make good gestures: shake_face, laugh, shy, wink. Clips install hot and persist in config.avatar.actions.

2. Questions. While the visitor talks, the engine asks one closed question per ~450 ms of new text: which of these clips fits right now, or none? While the avatar speaks it also asks does the reply text commit to a gesture? ("let me wave" → yes). Each option carries a one-line description so the model knows what the clip looks like.

3. Gate. A clip plays only when the pick is confident (default 0.6; 0.8 while text is still streaming), only once per cooldown (4 s), and only when the avatar is not mid-reply (speech frames keep priority — the clip waits for the next ask).

4. Fallback. No key, timeout, 429, or three failures in a row → the rules provider answers the same questions from keywords. Nothing ever blocks a turn.

Providers

Provider	Latency	Cost	Needs
rules	0 ms	free	nothing
jev	~100–500 ms	$0.042 / M input tokens	TypeSafe API key; an egress proxy where api.typesafe.ai answers 451

Configure in Console → Config → Instinct (config group instinct): enable, provider, key, optional proxy_url, thresholds. New sessions pick up changes.

Watching it work

The visitor page's DataChannel receives {"type":"instinct", clip, reason, provider, latency_ms, answers} for every decision; cored logs them as [instinct <session>]. The dsh plugin's mynah_action tool triggers the same clips by hand.