Evolving-Werewolf

Most AI Werewolf opponents share one fatal flaw: they never grow. Every match starts from the same playbook, and within a few rounds players learn to exploit the patterns. Evolving-Werewolf solves this by having the engine distill post-game review insights into a role-partitioned knowledge base after each match. At the start of the next game, those experiences are injected as skills into the new wave of player agents. Wolves develop sneakier kill strategies, seers sharpen their checks, and voting patterns grow more cunning. The more games played, the stronger the opponent collective becomes.

The engine is a complete nine-player Werewolf (3 Wolves / 1 Seer / 1 Witch / 1 Hunter / 3 Villagers), driven entirely by a state machine with no LLM gamemaster. This saves tokens and eliminates information distortion. Every role ability is fully implemented: wolf two-round discussion with majority vote, seer verification, witch heal and poison with first-night self-save, and the hunter last shot. The day flow covers sheriff election, directed speeches, recorded public voting with 1.5x sheriff weight, and tie-breaker revotes. Information isolation is strict — each role sees only its private channels, and incremental delivery pushes only what the player has not yet seen. A 60-second nudge, 180-second timeout skip, and a mutex lock keep long games from stalling; eliminated players have their info frozen so the post-game replay shows a causally clean decision trajectory.

This plugin suits three audiences: developers who want an evolving tabletop game embedded in DeepSeek Harness; researchers running multi-agent game-theory experiments who need observable cross-game strategy drift; and Werewolf enthusiasts who simply want to spar with a crew of personality-driven AI opponents that grow sharper with every session. The engine ships with random personality seeds for diverse play styles and an optional Zhipu AI vision and image-generation pipeline for ready-to-use avatars and character art. A human seat is assigned randomly, with all actions handled through a browser panel that competes fully alongside the AI agents.