Repository navigation
1.0.1
Minecraft CaYa Bot Panel v1.0.1
Stability and intelligence release after v1.0.0. This version rewrites the construction pipeline, fixes the “stuck after cancel” deadlock on Paper, and adds a full MCP / AI agent layer (Ollama in-game + Claude Code / Cursor / VS Code via /mcp).
Highlights
🏗 Construction engine rewrite (Faz 17 · closes #3)
- Chest stock ledger — nearby chests/barrels/shulkers are marked; stock counts as available; withdraw only when needed
- 3D-printer placement (default) — layer-by-layer serpentine; optional nearby-first
- Stall watchdog — recovery ladder instead of a blind hang
- Mid-build + final repair — damaged/wrong blocks re-queued; honest scaffold accounting
- Creative mode — material needs skipped; items provisioned for placement
- Reconnect resume — mid-build disconnect can continue after respawn
- Panel — place order, stored column, creative / stuck / repair / resume badges
🔓 Stop / Reset always frees the bot (Faz 17+ · closes #4)
Root cause was not only pathfinder: a hung build runner left TaskQueue.pump locked forever, so no later task (goto, stop, new build) could run — restart felt required.
Fix (three layers):
- Force-detach — cancelled runner that does not unwind within ~2s is abandoned; queue continues; panel logs a warning
- Bounded mineflayer ops — equip / place / dig / chest open-withdraw-deposit have hard timeouts and honor cancel
- Extended re-freeze — residual pathfinder goals cannot re-drive the bot after abort
Also: cancel-aware dig, companion pathing pause after Stop so follow does not immediately re-grab the bot.
🧠 MCP / AI agent system (Faz 18)
New panel tab MCP / AI with two modes sharing one tool registry (53 tools):
| Mode | How |
|---|---|
| Ollama (in-game) | Pick a local tool-capable model; enable agent per bot; talk from Agent Chat or in-game by name |
| MCP endpoint | http://127.0.0.1:3001/mcp — Claude Code, Cursor, VS Code Copilot, Windsurf, Gemini CLI, or stdio clients via mcp-remote |
Capabilities
- In-game replies when addressed (toggleable; off → panel Agent Chat only)
- Gather / mine / craft / drops with inventory targets (“have N of X”)
- Schematic-free creative building (geometric primitives) + library schematics
- Live build stage feedback (
get_build_status) - Protect / attack / flee / clear mobs (attack category off by default)
- Trust system — panel-managed; optional model-managed trust
- Persistent memory, autopilot goals, activity feed, persona & language
- Utility mode (default OFF) — own/permitted servers only: server commands, creative fly, utility dig; combat realism (D1–D8) never bypassed
Setup and client configs: see README → MCP / AI agent system.
# Claude Code example
claude mcp add --transport http caya-bot http://127.0.0.1:3001/mcp───
Commits (since v1.0.0)
┌─────────┬─────────────────────────────────────────────────────────────────────────────┐
│ Commit │ Change │
├─────────┼─────────────────────────────────────────────────────────────────────────────┤
│ c637145 │ feat(build): construction engine rewrite (issue #3) │
├─────────┼─────────────────────────────────────────────────────────────────────────────┤
│ 20e0a74 │ fix(build): force-detach hung runners — Stop/Reset frees the bot (Fixes #4) │
├─────────┼─────────────────────────────────────────────────────────────────────────────┤
│ 0f3233d │ feat(mcp): Ollama agent + /mcp endpoint, tools, trust, utility mode │
├─────────┼─────────────────────────────────────────────────────────────────────────────┤
│ 61db0ac │ docs: README MCP guide + roadmap (TODO) │
└─────────┴─────────────────────────────────────────────────────────────────────────────┘
───
Upgrade notes
- Pull / install v1.0.1 (or rebuild from master).
- Restart the panel (npm run dev or your desktop/portable build).
- MCP: enable the master switch under MCP / AI, pick an Ollama model and/or connect an MCP client. Utility mode stays off unless you turn it on for a permitted server.
- If a build was stuck on an older build: Stop or Reset work should free the bot within seconds; check the log for force-detached if something still misbehaves.
───
Notes / known limits
• Large schematic physics and some AI chat parse variants still benefit from Paper 1.21.x field checks.
• Small LLMs (≤8B) can mis-call tools; stronger tool-calling models are recommended for serious use.
• Offline (cracked) auth only; panel binds to localhost by default on purpose.
───
Full changelog: compare v1.0.0...v1.0.1 (v1.0.0...v1.0.1)
Docs: README (https://github.com/CaYatur/Minecraft-CaYa-Bot-Panel#readme)