DaemonCraft

DaemonCraft is a multi-agent Minecraft environment where persistent AI companions inhabit Mineflayer bodies, perceive and act through Hermes Agent, and coordinate with an embodied planning service.

Current architecture

• server/ — Purpur Minecraft server data, plugins, and backups
• agents/bot/ — Mineflayer body, HTTP/WebSocket API, movement, reflexes, and dashboards
• agents/agent_loop.py — body events, interoception, and gateway wake-up bridge
• agents/embodied-service/ — bounded embodied intent planning and dispatch
• agents/daemoncraft.py — cast and bot lifecycle management
• scripts/daemoncraft-ops.py — status, health, controller-mode, and operational probes
• docs/ — durable architecture, development, and operations documentation

DaemonCraft supports two deliberately separate modes:

• lab — interactive embodiment controlled from a Hermes client; no autonomous L4 turns
• autonomous — heartbeat-driven agent turns; enable only for bounded, observable experiments

Project authority

• GitHub Issues own actionable work, scope, dependencies, and acceptance criteria.
• Pull requests and commits carry implementation and verification evidence.
• docs/ contains durable architecture and operational knowledge.
• CHANGELOG.md contains release history.
• MEMORY.md, plans/, Lattice snapshots, and the retired local Kanban are historical evidence, not live trackers.

See AGENTS.md for the contributor and automation contract.

Operational safety

Do not start, restart, enable, or switch DaemonCraft services or autonomous mode without explicit maintainer instruction.

Read docs/OPERATIONS.md before operating the stack. A safe read-only preflight is:

./scripts/daemoncraft-ops.py status
./scripts/daemoncraft-ops.py health

The current recovery work is tracked in:

• #23 — Restore the known-good interactive lab embodiment baseline
• #24 — Diagnose autonomous-mode context divergence and action thrash
• #28 — Reconcile the runtime cognition contract

Development

• Use feature branches and pull requests; never commit directly to main.
• Search GitHub Issues and existing PRs before creating work.
• Preserve unrelated dirty changes.
• Add focused regression tests and report the exact verification commands in the PR.
• Never commit credentials or private runtime data; this repository is public.