Assistant Framework

Last updated: April 11, 2026

Assistant Framework

The Assistant Framework is the most complex part of the dLANDiscord backend, providing the intelligence and autonomy for all agent-based operations.

🧠 Core Methodology: ReAct & Planning

The framework implements a ReAct (Reason + Act) loop within a larger Adaptive Planning cycle.

  1. Thinking: The agent analyzes the task and its current context.
  2. Action: The agent decides to call a tool or delegate to another agent.
  3. Observation: The result of the action is fed back into the agent's context.
  4. Planning: The AdaptivePlanner evaluates if the current plan needs to be modified based on the observation (e.g., if a tool failed, it might replan with a different approach).

🗂️ Strategic Components

1. Adaptive Planner (adaptive_planner.go, planner.go)

Responsible for task decomposition. It breaks a high-level request (e.g., "Check all servers for security issues") into a tree of sub-tasks. It handles:

  • Fact Retention: Storing knowledge discovered during a session.
  • Hierarchical Delegation: Managing parent-child relationships between agents.
  • Dynamic Replanning: Using an LLM to recover from errors by adjusting the plan tree in real-time.

2. Zombie Contexts & Persistence

Asynchronous operations (like long-running background tasks) are handled via Zombie Contexts.

  • ZombieParentRegistry: Ensures that even if a parent session is cleaned up or the system restarts, the results of an async task can still be delivered to the correct user.
  • Race-Safe Delegation: Parent metadata is cloned before spawning background goroutines.
  • TTL & Cleanup: Stale async agents are cleaned up daily at 3:00 AM.

3. Mission Control (missioncontrol.go)

A centralized hub for real-time observation and administrative control. It provides:

  • SSE Neural Link: Live streaming of agent "thoughts" and tool execution.
  • Approval Pipeline: Manages tools and skills created by MrTool that require admin verification before going live.
  • Session Debugging: Fallback mechanisms (like reading JSONL logs directly) when the observability database is corrupted.

4. Heartbeat Loops (heartbeat.go)

These are proactive background routines that trigger specific agents for system maintenance:

  • Nikki Heartbeat: Regular check-ins for task board management.
  • Lenore Heartbeat: Infrastructure auditing and health checks.
  • CEO Heartbeat: Strategic oversight of all active projects.

🛠️ Tool System (mrtool.go)

The framework can execute over 100 different tools, ranging from system commands to API integrations.

  • Go Tools: Native Go functions.
  • Shell Tools: Bash/PowerShell scripts defined in .md files in assistant/prompts/tools/.
  • Worker Node Swarms: See AntFarm Swarm.

For detailed package documentation, see Assistant Core.