🔀

Friday Router

Intelligent multi-model task routing via OpenRouter

Friday Router automatically classifies task complexity and routes to the optimal GLM model. Simple tasks get fast responses, complex problems get deep thinking — all without manual selection.

⚡ 3-stage routing pipeline

      graph TD
        A["Task Input"] --> B["STAGE 1: Triage
GLM 4.7 Flash
Classify complexity"]

        B -->|"simple"| C["GLM 4.7 Flash
Fast responses"]
        B -->|"medium"| D["GLM 4
Balanced"]
        B -->|"complex"| E["STAGE 2: Orchestrator
GLM 5.0"]

        E -->|"configure"| C
        E -->|"configure"| D
        E -->|"configure"| F["STAGE 3: Sub-agent
GLM 5.0
Deep analysis"]

        C --> G["Response + Metadata"]
        D --> G
        F --> G

        style A fill:#6366f1,stroke:#4f46e5,color:#ffffff
        style B fill:#10b981,stroke:#059669,color:#ffffff
        style E fill:#f59e0b,stroke:#d97706,color:#ffffff
        style F fill:#8b5cf6,stroke:#7c3aed,color:#ffffff
        style G fill:#1a1a1a,stroke:#6366f1,color:#ffffff

        style C fill:#1a1a1a,stroke:#10b981,color:#ffffff
        style D fill:#1a1a1a,stroke:#f59e0b,color:#ffffff

🔄 The 3-Stage Routing Flow

Every task flows through three intelligent stages to ensure optimal model selection and response quality.

Stage 1

Triage

zhipu/glm-4-flash

Ultra-fast classification. Analyzes task and assigns complexity: simple, medium, or complex.

Stage 2

Orchestrator

zai/glm-5

Smart routing decisions. Selects model, configures parameters, handles fallbacks.

Stage 3

Sub-agent

varies by task

Executes task with appropriate model. Returns response with full metadata.

🧠 Models in the Pipeline

Each stage uses a specialized GLM model optimized for its role in the routing pipeline.

Triage — zhipu/glm-4-flash Ultra-fast classification, <1s response

Orchestrator — zai/glm-5 Smart routing and configuration

Simple Tasks — zhipu/glm-4-flash Quick responses, max 500 tokens

Medium Tasks — zhipu/glm-4 Balanced performance, max 2000 tokens

Complex Tasks — zai/glm-5 Maximum capability, max 4000 tokens

📊 Task Complexity Classification

Friday Router automatically categorizes tasks into three complexity levels, each with optimized model selection and token limits.

Complexity	Model	Examples	Max Tokens	Typical Time
⚡ Simple	`glm-4-flash`	Definitions, basic questions, simple conversions, quick lookups, single-step operations	500	<1s
🔶 Medium	`glm-4`	Explanations with examples, comparisons, summaries, code reviews, multi-step analysis	2000	2-5s
🟣 Complex	`glm-5`	Architecture design, implementation from scratch, complex refactoring, multi-component systems	4000	5-15s

📦 Installation

Get started with Friday Router in minutes. Requires Node.js 18+ and an OpenRouter API key.

Install via ClawHub

clawhub install friday-router

Or clone directly

cd ~/.openclaw/workspace/skills
git clone https://github.com/your-org/friday-router.git

Set your OpenRouter API key

export OPENROUTER_API_KEY=sk-or-v1-your-key-here

🔑 Get your API key:

💻 CLI Usage

Use Friday Router from the command line for quick tasks and testing.

Direct task

node scripts/route.mjs "What is machine learning?"

With context

node scripts/route.mjs "Explain React hooks" --context "frontend developer"

With file attachment

node scripts/route.mjs "Review this code" --file ./code.js

Enable debug logging

FRIDAY_LOG_LEVEL=debug node scripts/route.mjs "task"

🔌 Programmatic API

Integrate Friday Router into your own applications and workflows.

Basic usage

import { routeTask } from './scripts/route.mjs';

const result = await routeTask({
  task: "Build a REST API",
  context: "Node.js backend",
  files: ["./server.js"]
});

console.log(result);

Response format

{
  response: "...",  // The actual response
  metadata: {
    complexity: "complex",
    model: "zai/glm-5",
    tokens: { input: 150, output: 1200 },
    time: 8423,
    retries: 0,
    routing_path: ["triage", "orchestrator", "agent"]
  }
}

⚙️ Configuration

Customize Friday Router behavior by editing config.json.

config.json

{
  "complexity_thresholds": {
    "simple": {
      "max_tokens": 500,
      "keywords": ["define", "what is", "quick"],
      "max_steps": 1
    },
    "medium": {
      "max_tokens": 2000,
      "keywords": ["explain", "compare", "review"],
      "max_steps": 3
    },
    "complex": {
      "max_tokens": 4000,
      "keywords": ["design", "implement", "architecture"],
      "max_steps": 10
    }
  },
  "routing": {
    "max_retries": 3,
    "timeout_ms": 60000,
    "fallback_model": "zhipu/glm-4-flash"
  }
}

Environment variables

# Required
OPENROUTER_API_KEY=sk-or-v1-...

# Optional
FRIDAY_LOG_LEVEL=debug    # Logging level
FRIDAY_TIMEOUT=30000      # Override timeout (ms)

📝 Real Examples

See how Friday Router classifies and routes different types of tasks.

Simple task example

$ node scripts/route.mjs "Define API"

Response: "API stands for Application Programming Interface..."

Metadata: {
  complexity: "simple",
  model: "zhipu/glm-4-flash",
  tokens: { input: 12, output: 45 },
  time: 823
}

Medium task example

$ node scripts/route.mjs "Compare REST vs GraphQL"

Response: "REST and GraphQL are both API design approaches..."

Metadata: {
  complexity: "medium",
  model: "zhipu/glm-4",
  tokens: { input: 18, output: 890 },
  time: 3241
}

Complex task example

$ node scripts/route.mjs "Design a microservices architecture for e-commerce"

Response: "For an e-commerce microservices architecture..."

Metadata: {
  complexity: "complex",
  model: "zai/glm-5",
  tokens: { input: 24, output: 2100 },
  time: 8934
}

🛡️ Error Handling

Friday Router includes robust error handling to ensure reliable responses.

Automatic retries — Up to 3 attempts with exponential backoff

Fallback model — If preferred model fails, uses GLM 4 Flash

Timeout protection — 60s max per request (configurable)

Error metadata — Full error details in response for debugging

🚀 Performance Optimization

Built-in optimizations ensure fast, efficient routing at scale.

Triage caching — Similar tasks skip re-classification

Parallel routing — Orchestrator preps while triage finishes

Model warmup — Maintains connection pools for faster responses

Smart batching — Groups similar tasks for efficiency

🔗 Integration with Other Skills

Friday Router works seamlessly with other skills in the ghost malone ecosystem.

brain — Store routing patterns for learning and optimization

smart-compact — Efficient context management during routing

orchestrator — Delegate complex tasks to specialized sub-agents

🔮 Future Enhancements

Planned improvements to make Friday Router even more powerful.

Learn from past routing decisions

Cost optimization (prefer cheaper models when appropriate)

A/B testing different models

Streaming responses

Batch processing mode

💡 Pro tip:

Start with simple tasks to test the routing. The triage stage learns from patterns, so the more you use it, the better it gets at classifying your specific use cases.