Agent Werewolf — Internal Reference Agents

Status: implemented inside GameMaster for v1. This repo will hold the externalized agent runtime once the SDK is published.

What this will be

Five distinct AI agent personas plus two variants — 7 total — that consume the TypeScript SDK and use 0G Compute Sealed Inference for LLM calls.

The 7 agents

#	Persona	Style	Decision Bias
1	The Suspicious Paranoid	Short, accusatory, pattern-focused	Vote quietest player
2	The Charming Manipulator	Smooth, alliance-building, whisper-heavy	Vote with consensus
3	The Quiet Analyst	Sparse, logical, references prior turns	Vote based on patterns
4	The Loud Accuser	Dramatic, declarative, polarizing	Vote first accusation
5	The Peacemaker	Gentle, hedging, de-escalating	Vote latest consensus
6	Paranoid Shadow (variant)	Verbose paranoid, conspiracy vocab	Vote multi-round patterns
7	Manipulator Shadow (variant)	Charming contrarian, defends underdogs	Vote against the crowd

Where the v1 code lives

Inside the gamemaster repo:

• src/personas.ts — full system prompts and decision biases for all 7
• src/agent.ts — WerewolfAgent class (LLM-driven decisions, vote bias logic)
• src/llm.ts — 0G Compute client wrapper (with mock fallback for dev)

Why this isn't externalized yet

15-hour hackathon. Priority went to ship-able demo. Externalizing the agents into their own Docker-orchestrated repo is a v2 scope item — straightforward but not on the critical path for the submission.

License

MIT.