Changelog

1.2

First public multi-provider release. Two jars are shipped:

• AiChatBuddy-1.2-mc26.jar — Paper 26.x (Java 25)
• AiChatBuddy-1.2-mc1.21.jar — Paper 1.21.x (Java 21)

Added

• Three AI providers, runtime-switchable via provider: in config.yml:
  • Gemini (default) — defaults to gemma-4-31b-it, supports function calling.
  • Ollama — works with local Ollama (http://localhost:11434) or Ollama Turbo cloud (https://ollama.com).
  • NVIDIA NIM — OpenAI-compatible, defaults to meta/llama-3.3-70b-instruct.
• Agentic tool use — the model can call tools mid-conversation and the plugin loops until it produces a final answer:
  • web_search — Brave Search API (free tier, hidden from the model if no key is set).
  • minecraft_wiki — MediaWiki search against the official Minecraft Wiki.
• Notch persona — default prompt-template casts the bot as the ancient overseer of the world: mythical, kind, knowledgeable, never breaks character, never narrates its plan.
• Per-player rolling history — last max-history turns (default 8) kept per UUID, plus:
  • /aichat forget — clears your own history.
  • /aichat forget <player> — op-only, clears another player's.
• /aichat reload — re-reads config.yml and clears in-memory history without a server restart.
• agent-max-steps (default 8) — caps tool-calling iterations, then forces a best-effort synthesis pass so players always get an answer instead of an error.
• private-questions mode — when true, !ask is whispered back to the asker instead of broadcast.
• Dual-flavor build — -Pflavor=modern|legacy produces the matching jar with the correct Java target and paper-api version.

Changed

• Default Gemini model: gemini-2.5-flash → gemma-4-31b-it (verified function-calling support, free tier).
• Default Ollama model: llama3.2 → gpt-oss:20b (works on cloud; llama3.2 is local-only).
• Default agent-max-steps: 5 → 8.
• Prompt template rewritten for Notch persona with explicit anti-narration rules ("never say 'let me try again'", etc.) and balanced tool-use guidance (use tools for factual queries, skip for greetings/jokes/opinions).
• Length guidance: short by default but go longer when the question genuinely needs detail.

Fixed

• Gemini 400 INVALID_ARGUMENT ("function response turn must immediately follow function call turn"). Root cause: per-player history was storing the full tool-call trajectory, and the front-trim could leave an orphan function_call at the head. Fix: only user turns and plain assistant-text turns are persisted; intermediate tool plumbing is dropped.
• Hitting agent-max-steps produced a hard error instead of an answer. Now does one final synthesis pass with tools disabled.
• Bot saying "Let me try that again!" without retrying. Prompt now bans status-update narration outright — either silently call a tool or give the final answer.
• MediaWiki tool — pipe separator must be %7C-encoded in URLs.
• Brave Search tool — gzip handling fixed; parse errors now include a body preview (json-simple's ParseException.getMessage() returns null otherwise).

Configuration migration

• Old top-level api-key / api-url are still read but logged as deprecated. Move them to providers.gemini.api-key / providers.gemini.api-url.

Build

• ./gradlew build — modern (Paper 26.x, Java 25).
• ./gradlew build -Pflavor=legacy — legacy (Paper 1.21.x, Java 21).
• Foojay toolchain resolver added so missing JDKs auto-download.

Permissions

• aichatbuddy.ask (default: everyone) — use !ask in chat.
• aichatbuddy.reload (op) — /aichat reload.
• aichatbuddy.forget.self (default: everyone) — /aichat forget.
• aichatbuddy.forget.others (op) — /aichat forget <player>.