Jarvis: AI Butler

Jarvis now runs on Fabric and NeoForge as well as Paper. One release, three files — pick the one for your server. Same butler, same commands, same config file on all three.

Upgrading from 0.16 on Paper

• The plugin file is now named jarvis-paper-0.17.0.jar. Delete the old Jarvis-0.16.0.jar from plugins/ so you do not load two copies. Your config and data carry over unchanged.
• The jar is much larger (~86 MB, was ~1 MB). It now carries the speech engine's native libraries for Linux, Windows and macOS, so voice works with nothing else installed. If you never turn voice on, they are never loaded.
• build.fallback-material is retired; blocks the server does not recognise are now skipped and named in the log instead of becoming dirt.

New

• Fabric and NeoForge. On the mods the butler is a server-side fake player with his own pathfinding — no Citizens needed. Admin means operator. Schematics paste on every platform; saving and rotating them needs WorldEdit, so those two are Paper-only. Config lives in config/jarvis/.
• AI setup from the bell menu. Admin → AI setup: switch providers on and off, enter keys (typed in chat but kept out of the chat log), pick an Ollama model from a list, and test the connection. /jarvis ai … does the same from the console. Saving keeps the comments in your config.yml.
• Better freeform builds. The model describes a building as shapes — walls, floors, doors, beds, pitched roofs — so walls come out whole, doors have both halves and roofs close their gables. Houses come furnished: bed, chest, crafting table, light. Doors and windows a block off the wall are snapped into it. Building → Custom build in the bell menu.
• Voice runs inside the server on all three platforms, through Simple Voice Chat. Speech recognition (whisper.cpp) and his voice (Piper) run in the game server — no separate service. About 200 MB of voice models are downloaded the first time you turn voice on. Admin → Voice setup, or /jarvis voice enable. /jarvis voice shows every link of the chain, and /jarvis voice bench finds the fastest thread count for your machine.

Changed

• He is summoned two blocks away, facing you, instead of on your head.
• Gear he was issued stays his: dismissing him hands back everything except his kit, and his inventory screen won't let it be taken.
• Voice recognition is faster: it only processes the clip you spoke.
• Large build plans no longer stop half-way through.

Fixed

• "Failed to generate build plan" when the model's reply was cut short — the complete part of the plan is now salvaged and built.
• Jarvis failed to start on a server without GraalJS.

Full technical changelog: https://github.com/iamgadgetman/jarvis/blob/main/CHANGELOG.md