intelligent-router_skill

Overview

This skill provides intelligent model routing for sub-agent task delegation. It selects the optimal model based on task complexity, capability needs, and cost, routing simple work to cheaper models and preserving premium models for high-stakes tasks. The router supports tiered classification, confidence scoring, fallback chains, and agentic task detection to reduce cost and failure rates.

How this skill works

The router analyzes task text with a 15-dimension weighted scoring system and computes a confidence score via a smooth S-curve. It classifies tasks into SIMPLE, MEDIUM, COMPLEX, REASONING, or CRITICAL tiers and picks models configured in your config.json. Automatic fallback chains and retry logic handle failures; agentic patterns bump multi-step tasks to higher tiers. Cost estimates and coding-specific flows guide whether to delegate or run the task on a single premium model.

When to use it

• When spawning sub-agents or delegating tasks across multiple models
• When you need to balance cost vs. quality for many small tasks
• Before running multi-step or tool-using agent workflows
• When you want automatic retries and fallback behavior on model failure
• To profile and estimate model costs before execution

Best practices

• Configure all available models with accurate costs, capabilities, and context windows in config.json
• Run the classifier or score CLI before spawning sub-agents to check tier and confidence
• Use the REASONING tier only for formal proofs or step-by-step logic with high confidence thresholds
• For coding: prefer MEDIUM + SIMPLE QA for small patches and COMPLEX/CRITICAL for multi-file or architecture work
• Keep fallback chains short (max 3 attempts) and document model failure patterns for future routing improvement

Example use cases

• Route monitoring and summarization tasks to SIMPLE-tier models to save cost
• Classify and route code fixes: MEDIUM model for single-file fixes, COMPLEX for multi-file refactors
• Detect agentic tasks (deploy, test, build) and escalate to appropriate tiers automatically
• Estimate costs for a planned workload and compare coder+QA cost vs premium model cost
• Use REASONING tier for theorem proofs or formal verification with confidence >= 0.97