Changelog

v1.4.2

Security & Permission Model Changes

• execute_command no longer executes with elevated server-side privileges — Vanilla commands suggested by the AI are no longer executed directly on the server with OP permissions. Instead, they are sent to the requesting player's client, which automatically opens the chat screen with the command pre-filled. Whether the command is sent, and with what permission level, is entirely up to the player, removing the risk of "remote backend control of privileged command execution." The server-side command blacklist has been removed accordingly, since the server no longer executes commands directly.
• New standalone AI Permission Settings screen — Split out from AI Chat Settings into K → AI Permission Settings, centralizing the following permission-related options:
  • Allow AI to Use Vanilla Commands (vanilla_commands_enabled, on by default): when off, the system prompt no longer includes execute_command instructions, so the AI can only use the mod's built-in features; even if the AI still tries to generate that instruction, it is rejected outright at runtime.
  • Require Confirmation Before Execution (confirm_before_execute_enabled, off by default): when on, before the AI uses a mod-specific feature (placing blocks, building blueprints, terrain queries, web search, web scraping, etc.) it first sends a [Yes]/[No] confirmation message in chat. Multiple actions within the same round are merged into a single batch confirmation, and unconfirmed requests are automatically canceled after 60 seconds.
  • Max tool-call rounds (the max_tool_rounds setting, moved here from the chat settings screen).
• New /aiconfirm command — Handles [Yes]/[No] button clicks in chat for pending AI action requests. Once confirmed, the conversation automatically resumes, covering all tool types (web search, web scraping, terrain queries, in-game actions), preventing the conversation from stalling or actions from being silently dropped after confirmation.

Improvements

• Removed hard block-count caps for fill/clear/terrain queries — Replaced with soft prompt-level guidance (fill/clear recommended to stay under ~10,000 blocks, terrain queries under ~30,000 blocks in volume). Exceeding these is no longer rejected with an error; instead the AI is prompted to split the operation into multiple calls, reducing unnecessary failures.
• Removed the teleport tool — The AI-triggerable teleport instruction has been removed, simplifying the set of available actions.
• Refactored the multi-round tool loop — Introduced a unified finishing mechanism (ToolLoopFinisher) ensuring the confirmation flow and network calls (web search, web scraping, resuming the AI) all run on separate threads instead of blocking the server's main tick thread.

Documentation

• Synced the README, user manuals (Chinese & English), and in-game manuals with the latest behavior: updated the execute_command permission model description, added documentation for the new permission settings screen, the /aiconfirm command, and new config options; removed outdated teleport references; bumped the version number to 1.4.2.