L2 Planner Architecture

Overview

The L2 planner is a hybrid LLM-A* planning system that sits between the deterministic L0/L1 layers and human escalation (L3). It combines creative plan generation from an LLM agent (Ava) with rigorous A* validation and optimization.

Layer Model

L0 (Deterministic Rules) → L1 (A* Search) → L2 (LLM-A* Hybrid) → L3 (Human)

• L0: Pattern-matching rule engine. Fast, deterministic, zero-cost. Handles ~80% of cases.
• L1: A* search with HTN decomposition. Finds optimal plans within budget. Handles ~15%.
• L2: LLM proposes creative plans, A* validates and optimizes. Handles novel situations.
• L3: Human-in-the-loop for cases where L2 confidence is too low.

L2 Hybrid Planning Flow

1. Routing: L2Router intercepts L0/L1 failures or low-confidence results
2. Proposal: A2AProposer sends structured prompt to Ava (LLM) via A2A protocol
3. Validation: AStarValidator checks each candidate plan against A* simulation
4. Optimization: A* attempts to find a lower-cost alternative
5. Confidence: ConfidenceScorer evaluates plan quality (feasibility, goal alignment, cost, constraints)
6. Escalation: EscalationTrigger routes low-confidence plans to L3

Learning Flywheel

Every successful L2 plan is fed into the learning flywheel:

1. RuleExtractor extracts generalizable conditions from the plan
2. PlanConverter creates a LearnedRule in the RuleRegistry
3. On subsequent requests, PatternMatcher checks learned rules first
4. After enough successful executions, RuleMigration promotes the rule to L0
5. Result: escalation rate decreases over time as the system learns

Confidence Scoring

Confidence is a weighted composite of:

• Feasibility (35%): Can actions execute in order?
• Goal Alignment (30%): Does the final state satisfy the goal?
• Cost Efficiency (15%): Is the cost reasonable?
• Constraint Satisfaction (20%): Are all constraints met?

Configuration

See src/config/routing-config.yaml for tunable thresholds.

Key Files

File	Purpose
src/planner/l2-router.ts	Top-level L0→L1→L2→L3 routing
src/planner/hybrid-planner.ts	LLM-A* hybrid planning engine
src/planner/a2a-proposer.ts	LLM candidate generation via A2A
src/planner/astar-validator.ts	A* plan validation and optimization
src/planner/confidence-scorer.ts	Plan quality scoring
src/planner/escalation-trigger.ts	L3 escalation decisions
src/planner/dispatcher.ts	Full pipeline dispatcher
src/learning/plan-converter.ts	Plan → rule conversion
src/learning/rule-registry.ts	Learned rule storage
src/learning/rule-migration.ts	L2 → L0 rule promotion
src/monitoring/l2-metrics.ts	Invocation/success telemetry
src/monitoring/escalation-tracker.ts	Escalation trend tracking