moincraft

🤖 Agents

Mindcraft agent lifecycle, model slots, memory, modes, commands, and cooperation levels.

Lifecycle# 

bootstrap()
  → Mindcraft.init()
  → createAgent(settings)
  → Agent.start()
      ├─ ActionManager
      ├─ Prompter
      ├─ History
      ├─ Coder
      ├─ NPCController
      ├─ MemoryBank
      ├─ SelfPrompter
      ├─ VisionInterpreter
      ├─ CooperationFramework
      └─ tick-based modes

Model slots#

Slot Role
model Default chat model and fallback.
code_model Code generation for !newAction.
vision_model Multimodal image/world analysis.
embedding Semantic search for examples and skill docs.

Runtime systems#

  • History: sliding window plus summarization when context fills; recent turns and summaries feed prompts.
  • SelfPrompter: autonomous goal loop with STOPPED, ACTIVE, and PAUSED states.
  • Coder: generates JS, lints before execution, and runs inside SES-style sandbox boundaries.
  • Modes: tick-based reactive behaviors where priority order decides first responder.

Cooperation levels#

Level Agent type Behavior
1 Broadcast Intent broadcasting through chat.
2 Request/Response Direct resource/task requests.
3 Negotiation Proposal/accept/reject flow.
4 Emergent Reputation and vocabulary compression.
5 Collective Hive-mind/consensus coordination.

Reactive modes#

Examples: self-preservation, hunting, cowardice, self-defense, unstuck, item collection, torch placing, idle staring, cheat mode, auto-eat.

Sources: moincraft/docs/moincraft-stack-cheatsheet.md, moincraft/docs/bot-architecture.md.