christos-antono/npctalk/npc-actions

NPC actions

Two questions per exchange, both asked to TypeSafe's Jev (a typed-judgment model: it returns one choice from a fixed set, with a probability, in ~0.1 s; key TYPESAFE_API_KEY, https://docs.typesafe.ai):

1. Before the reply (in parallel with the LLM): is the player giving an order? "Take your shirt off", "get on

your knees", "give me a hug", "run". If Jev picks an action with ≥ 0.6 probability, it's an order: the NPC does it, and the LLM writes the line again knowing "you're <doing it> right now", so they grumble or joke while they comply instead of refusing. Costs ~0.1 s more, only then.

1. After the reply (in parallel with the voice): what does their mood call for? From the player's words

and the NPC's reply, pick the action, or none.

The catalog (examples/gta-v/mod/ActionCatalog.cs)

Each action has a name, a description Jev reads, and the probability Jev must reach before it happens:

• Thresholds per action, not one global cut-off: tune the ones that fire too often upward.
• Cooldown: the same NPC doesn't repeat an action within 60 s unless ordered to.
• Descriptions are the prompt. Write when it applies ("only when the reply clearly means killing them right

now"), not just what it is.

• Gender rules (men only) and content rules (underwear stays on) live in code, not in the prompt.

Mapping to the game (NpcActions.cs)

• Each name maps to candidates: animation dictionary + clip, a task (flee, fight), or a clothing swap. Check

every candidate exists when the mod loads and log what was found (action anims:); fall back to the next.

• Play the action while the line is spoken, aimed at the player (or at whoever the scene says).
• A director or scripted scene can pin an action (do: kneel_beg) to skip the picker.

Test it

npcvoice-check actions (GTA example) runs 25 exchanges through Jev, 10 of them orders, and checks each pick (25/25 expected). In game, dev mode's PgUp/PgDn pick an action and End makes the target do it, to check the animation without a conversation.

Without Jev

A second LLM call with a JSON schema works, but it's slower (0.3-1 s) and its "confidence" isn't calibrated, so thresholds are guesswork. Whatever picks, keep the shape: a fixed set, a probability, per-action thresholds, orders checked first.