planning-goal（规划目标）

Goal-Oriented Action Planning (GOAP)

Dynamic planning system using A* search to find optimal action sequences for complex objectives

Quick Start

# Define goal state and current state
Current: {code_written: true, tests_written: false, deployed: false}
Goal: {deployed: true, monitoring: true}

# GOAP generates optimal plan:
1. write_tests -> tests_written: true
2. run_tests -> tests_passed: true
3. build_application -> built: true
4. deploy_application -> deployed: true
5. setup_monitoring -> monitoring: true

When to Use

• Complex multi-step tasks with dependencies requiring optimal ordering
• High-level goals needing systematic breakdown into concrete actions
• Deployment workflows with many prerequisites
• Refactoring projects requiring incremental, safe transformations
• Any task where conditions must be met before actions can execute

Prerequisites

• Clear definition of current state (what is true now)
• Clear definition of goal state (what should be true)
• Available actions with known preconditions and effects

Core Concepts

GOAP Algorithm

GOAP uses A* pathfinding through state space:

1. State Space: All possible combinations of world facts
2. Actions: Transforms with preconditions and effects
3. Heuristic: Estimated cost to reach goal from current state
4. Optimal Path: Lowest-cost action sequence achieving goal

Action Definition

Action: action_name
  Preconditions: {condition1: true, condition2: value}
  Effects: {new_condition: true, changed_value: new_value}
  Cost: numeric_value
  Execution: llm|code|hybrid
  Fallback: alternative_action

Execution Modes

Implementation Pattern

interface WorldState {
  [key: string]: boolean | string | number;
}

interface Action {
  name: string;
  preconditions: Partial<WorldState>;
  effects: Partial<WorldState>;
  cost: number;
  execution: 'llm' | 'code' | 'hybrid';
  tools?: string[];
  fallback?: string;
}

interface Plan {
  actions: Action[];
  totalCost: number;
  estimatedTime: string;
}

function generatePlan(
  currentState: WorldState,
  goalState: WorldState,
  availableActions: Action[]
): Plan {
  // A* search through state space
  // Returns optimal action sequence
}

Configuration

goap_config:
  planning:
    algorithm: a_star
    max_depth: 50
    timeout_ms: 5000

  execution:
    mode: adaptive  # focused | closed | open
    parallel_actions: true
    replan_on_failure: true

  monitoring:
    ooda_loop: true
    observe_interval_ms: 1000

  cost_weights:
    time: 1.0
    risk: 2.0
    resource: 1.5

Usage Examples

Example 1: Software Deployment

current_state:
  code_written: true
  tests_written: false
  tests_passed: false
  built: false
  deployed: false
  monitoring: false

goal_state:
  deployed: true
  monitoring: true

available_actions:
  - name: write_tests
    preconditions: {code_written: true}
    effects: {tests_written: true}
    cost: 3

  - name: run_tests
    preconditions: {tests_written: true}
    effects: {tests_passed: true}
    cost: 1

  - name: build_application
    preconditions: {tests_passed: true}
    effects: {built: true}
    cost: 2

  - name: deploy_application
    preconditions: {built: true}
    effects: {deployed: true}
    cost: 2

  - name: setup_monitoring
    preconditions: {deployed: true}
    effects: {monitoring: true}
    cost: 1

# Generated Plan (cost: 9)
plan:
  1. write_tests
  2. run_tests
  3. build_application
  4. deploy_application
  5. setup_monitoring

Example 2: Complex Refactoring

current_state:
  legacy_code: true
  documented: false
  tested: false
  refactored: false

goal_state:
  refactored: true
  tested: true
  documented: true

generated_plan:
  1. analyze_codebase:
      effects: {understood: true}

  2. write_tests_for_legacy:
      requires: understood
      effects: {tested: true}

  3. document_current_behavior:
      requires: understood
      effects: {documented: true}

  4. plan_refactoring:
      requires: [documented, tested]
      effects: {plan_ready: true}

  5. execute_refactoring:
      requires: plan_ready
      effects: {refactored: true}

  6. verify_tests_pass:
      requires: refactored
      validates: goal_achieved

Example 3: OODA Loop Monitoring

// Observe-Orient-Decide-Act loop during execution
async function executeWithOODA(plan: Plan): Promise<Result> {
  for (const action of plan.actions) {
    // OBSERVE: Check current state
    const currentState = await observeState();

    // ORIENT: Analyze deviations
    const deviation = analyzeDeviation(currentState, expectedState);

    // DECIDE: Replan if needed
    if (deviation.significant) {
      const newPlan = await replan(currentState, goalState);
      return executeWithOODA(newPlan);
    }

    // ACT: Execute action
    await executeAction(action);
  }
}

Execution Checklist

• Define current state completely
• Define goal state with all required conditions
• Inventory available actions with preconditions/effects
• Calculate action costs realistically
• Generate plan using A* search
• Review plan for feasibility
• Execute with OODA loop monitoring
• Handle failures with adaptive replanning
• Verify goal state achieved

Best Practices

• Atomic Actions: Each action should have one clear purpose
• Explicit Preconditions: All requirements must be verifiable
• Predictable Effects: Action outcomes should be consistent
• Realistic Costs: Use costs to guide optimal path selection
• Replan Early: Detect failures quickly and adapt
• Parallel Where Possible: Execute independent actions concurrently

Error Handling

Plan Generation Failures

// No valid path exists
if (!plan) {
  // Analyze which preconditions cannot be satisfied
  const unsatisfiable = findUnsatisfiablePreconditions(goalState);
  console.error(`Cannot reach goal: missing ${unsatisfiable}`);

  // Suggest partial goals that ARE achievable
  const partialGoals = suggestAchievableSubsets(goalState);
}

Execution Failures

// Action failed during execution
if (actionResult.failed) {
  // Check if alternative action available
  if (action.fallback) {
    await executeAction(action.fallback);
  } else {
    // Replan from current state
    const newPlan = await replan(currentState, goalState);
  }
}

Metrics & Success Criteria

Integration Points

MCP Tools

// Orchestrate GOAP plan across swarm
  task: "execute_goap_plan",
  strategy: "adaptive",
  priority: "high"
});

// Store successful patterns
  action: "store",
  namespace: "goap-patterns",
  key: "deployment_plan_v1",
  value: JSON.stringify(successfulPlan)
});

Hooks

# Pre-task: Initialize GOAP session

# Post-task: Store learned patterns

Related Skills

• planning-code-goal - SPARC-enhanced code planning
• sparc-workflow - Structured development
• agent-orchestration - Swarm coordination

References

• GOAP in Game AI
• A* Search Algorithm
• OODA Loop

Version History

• 1.0.0 (2026-01-02): Initial release - converted from goal-planner agent