Kai 2.0

A persistent, PvP/survival AI agent for Paper that drives a FakePlayer bot to behave like a real player. Kai runs fully in-process — there is no external backend (Kai 1.0's OpenRouter LLM brain has been removed).

Requirements

• Paper 26.1.2 (Maven artifact io.papermc.paper:paper-api:26.1.2.build.67-stable)
• Java 25
• A FakePlayer plugin installed on the server. The ported bot controller (NMSBot) talks to the plugin named FakePlayerPlugin via reflection (getFppApi().spawnBot(Location, Player, String)), which is what Kai 1.0 used. Declared as softdepend.
• (Optional) LuckPerms — NMSBot preloads the bot's LuckPerms user if present (cosmetic, guarded).

Runtime note: a 26.1.x-compatible FakePlayer build is required for the bot to actually spawn. Kai itself is version-independent — it compiles against paper-api + the public Bukkit/Paper API only.

Architecture

Kai 2.0 keeps the deterministic, local subsystems from Kai 1.0 and drops everything that depended on an external service or was outside the PvP/survival scope (OpenRouter gateway, chat moderation, shop, personality/tease, grief logging, bridge writer).

Package	System	Responsibility
JonaSMP_AI	main	Plugin entry; owns the shared ChunkRadar (getInstance().getChunkRadar()) and the bot
watcher.AIPlayerBot	brain	Per-tick orchestrator: combat → goals → equip/enchant → inventory/craft → pickup
watcher.BotGoalPlanner	goals	Deterministic GOAL→STEP planner (gather/mine/build/survive/follow)
watcher.BotCombatManager	pvp	Target selection, kiting, crit timing, mace-smash, shield-break, W-tap, auto-eat
watcher.BotPathfinder	pathing	A* navigation
watcher.NMSBot	bot	Spawns & controls the FakePlayer (FPP reflection + public Paper API)
watcher.BotMemory(Storage)	memory	Persistent player/threat/resource knowledge with TTL
radar.ChunkRadar (+ChunkCache/ChunkFileReader/DeltaTracker)	radar	Tiered chunk awareness: live deltas → RAM cache → loaded chunks
watcher.BotAutoEquipper/BotAutoEnchanter/CraftingPlanner/ContainerScanner/BurrowManager/WorldProtection	support	Gear, enchanting, crafting, container scanning, combat-burrow, anti-grief guards

Key changes vs. Kai 1.0

• LLM removed. Kai 1.0 asked an OpenRouter model for one action roughly every second — the cause of the sluggish "one decision per tick" feel. Kai 2.0's brain is the deterministic BotGoalPlanner
  • BotCombatManager, which were already the primary drivers in 1.0; the LLM was only a fallback.
• Control loop reactivity. The main AIPlayerBot tick task ran on a 20-tick (1 Hz) period in 1.0, so combat/movement only updated once per second. It now runs every tick (runTaskTimer(..., 20L, 1L)), with expensive perception/equip work kept on its own cadences.
• Radar thread-safety. The radar's chunk read ran on an async task in 1.0 while calling World#getChunkAt / Chunk#getBlock (not thread-safe). It now runs on the main thread and reads at most one uncached chunk per cycle to keep per-tick cost flat.

Build

mvn clean package

The plugin jar is produced at target/Kai-2.0.0.jar.

Commands

/kai <spawn|despawn|come|say> (permission kai.command, default OP)

• spawn [name] — create Kai at your location via FakePlayer (default name Kai)
• despawn — remove Kai
• come — order Kai to walk to you
• say <text> — issue a natural-language order routed to the deterministic goal planner