Goal & Planning¶
Goal decomposition, navigation, and orchestration -- the components that turn high-level objectives into executable plans with safety-bounded subtasks.
Overview¶
The goal engine is the strategic brain of G6. Given a GoalInput (a frozen Pydantic model with nested subtasks and resource bounds), GoalDecomposer builds a SearchTree of subgoals, each annotated with resource constraints inherited from the parent. The engine exposes 325 MCP tools across 13 sub-packages covering goal lifecycle, mental models, metacognition, orchestration, persistence, verification, and EML (Epistemic Meta-Learning).
Supporting the engine are five orchestration components. solver handles constraint satisfaction for resource allocation. navigator provides pathfinding through the goal search tree. guide offers heuristic advice for goal refinement. hat_orchestrator manages hierarchical agent teams, and recursive_architect handles recursive goal decomposition when subtasks themselves require decomposition.
The cluster enforces a strict resource budget: every subtask inherits a ResourceBoundsSchema from its parent, and the sum of child budgets must not exceed the parent's allocation. This prevents unbounded computation even in deep recursive decompositions.
Components¶
| Component | Description | MCP Tools |
|---|---|---|
| goal_engine | Goal decomposition, SearchTree construction | 325 |
| solver | Constraint satisfaction for resource allocation | -- |
| navigator | Search tree pathfinding and traversal | -- |
| guide | Heuristic goal refinement advice | -- |
| hat_orchestrator | Hierarchical agent team management | -- |
| recursive_architect | Recursive goal decomposition | -- |
Architecture¶
graph TD
GE[goal_engine] --> SOLVER[solver]
GE --> NAV[navigator]
GE --> GUIDE[guide]
GE --> HAT[hat_orchestrator]
GE --> RA[recursive_architect]
RA -->|deep subtasks| GE
HAT --> AGENTS[Agents & LLM cluster]
GE --> CTX[Context & Retrieval cluster]
GE --> SAFE[Safety & Alignment cluster] Key Patterns¶
Frozen Goal Inputs. GoalInput and ResourceBoundsSchema are frozen Pydantic models. Once a goal is submitted, its specification is immutable. This simplifies reasoning about concurrent goal execution and prevents mid-flight mutation.
MCP Sub-Package Organisation. The goal engine's 325 tools are split across 13 sub-packages (Round 1: goal_engine_mcp, mental_models_mcp, metacognition_mcp, orchestration_mcp; Round 2: persistence_mcp, verifier_mcp, eml_mcp; plus six more). Each sub-package has its own schema, server, and test suite, keeping individual modules under 300 lines.
Resource Inheritance. When GoalDecomposer splits a goal into subtasks, each child receives a ResourceBoundsSchema derived from the parent. The decomposer validates that child bounds sum to at most the parent's bounds, providing a compile-time-like check on resource consumption.
Related Clusters¶
- Safety & Alignment -- goals are verified against alignment specs
- Context & Retrieval -- goal decomposition uses retrieved context
- Agents & LLM -- agents execute the leaf tasks in a goal tree
- Experience & Autonomy -- experience loops refine goal strategies