AGENT SKILL

llm-orchestration

SWAL Unified LLM Orchestration - Combines resource monitoring, multi-key Gemini management, and OpenRouter routing. Intelligent provider selection for code, fast, deep, and reasoning tasks. Version 1.0 | Updated 2026-04-23

License 许可
MIT
所属仓库
iberi22/skills
最近更新
2026-10-12
Stars
★ 0
Forks
0

01 前言 Front Matter

name
llm-orchestration
metadata
{ "author": "swal", "unified": "2026-04-23", "updated": "2026-04-23", "version": "1.0.0", "openclaw": { "autoLoad": false } }
description
SWAL Unified LLM Orchestration - Combines resource monitoring, multi-key Gemini management, and OpenRouter routing. Intelligent provider selection for code, fast, deep, and reasoning tasks. Version 1.0 | Updated 2026-04-23

03 SKILL.md 正文

LLM Orchestration — Unified SWAL LLM Provider System

⚡ Intelligent Routing Across All LLM Providers

Combines llm-resource-monitor + gemini-key-pool + openrouter into a single unified orchestration system.


🎯 PURPOSE

Unified monitoring and intelligent orchestration of all LLM providers:

  1. Scans provider capabilities (context, limits, cost)
  2. Tracks usage in real-time (RPM, daily, tokens)
  3. Manages multi-key failover (Gemini-specific)
  4. Routes tasks to optimal provider based on requirements
  5. Orchestrates intelligent fallback on rate limits

📊 UNIFIED PROVIDER STATUS

Provider Context Daily Limit RPM Cost Best For
🟢 Groq 128K 14,400 30 Free Fast tasks, speed
🔵 Gemini #1-4 1M 1,500/key 15 Free Deep reasoning
🟡 DeepSeek 163K 10,000 60 $0.81 Reasoning, cheap
🟠 OpenRouter 262K 200 20 $0.35 Model variety
🟣 MiniMax ∞ ∞ ∞ Unlimited Coding, fallback
🔵 Codex (GPT-5.4) 258K→1M 30-150/5h varies Plus Heavy coding

🚀 USAGE

Dashboard (Full Status)

node scripts/llm-resource-monitor.js status

Best Provider for Task

node scripts/llm-resource-monitor.js best code
node scripts/llm-resource-monitor.js best fast
node scripts/llm-resource-monitor.js best deep
node scripts/llm-resource-monitor.js best cheap

Intelligent Orchestration

node scripts/llm-resource-monitor.js orchestrate code 50000

Record Usage

node scripts/llm-resource-monitor.js record groq 500

🔑 GEMINI KEY POOL (Multi-Key Management)

⚠️ CRITICAL: Each key must be from a DIFFERENT Google AI Studio project.

Environment Variables

export GOOGLE_API_KEY="key-from-project-1"
export GOOGLE_API_KEY_2="key-from-project-2"
export GOOGLE_API_KEY_3="key-from-project-3"
export GOOGLE_API_KEY_4="key-from-project-4"

Check Status

node scripts/gemini-key-pool.js status

Make a Request

node scripts/gemini-key-pool.js call --prompt "..." --model gemini-3.1-pro

Deep Research

node scripts/gemini-key-pool.js research "complex query"

Key Status Display

Status Meaning Action
🟢 OK <50% used Normal use
🟡 MEDIUM 50-80% Consider switching
🔴 HIGH 80-100% Switch provider
⏸️ EXHAUSTED Rate limited Wait 1 hour

Capacity Calculation

Keys Requests/Day Use Case
1 1,500 Light usage
2 3,000 Moderate usage
4 6,000 Heavy usage (recommended)

🌐 OPENROUTER (Free Models)

Free Models Available

Model Context Best For
DeepSeek R1 128K Reasoning, debugging
Qwen3.5-9B 32K Light tasks, quick fixes
Elephant Alpha 256K General intelligence
Qwen3.6 Plus Preview 1M Complex analysis

Setup

# Set API key
export OPENROUTER_API_KEY="sk-or-v1-your-key"

# Configure in OpenCode
opencode run --model deepseek/deepseek-r1 "task"

🎯 TASK TYPE ROUTING

Task Primary Secondary Notes
code MiniMax Groq OpenCode for coding
fast Groq MiniMax Groq for speed
deep / research Gemini #1-4 DeepSeek 1M context
reasoning DeepSeek Groq Reasoning tasks
cheap DeepSeek OpenRouter Budget-friendly
heavy Codex (GPT-5.4) MiniMax Plus account

🔍 SCANNING LOGIC

The scanner evaluates:

  1. Context Match — Can handle tokens needed?
  2. Daily Limit — Enough remaining quota?
  3. RPM — Under rate limit currently?
  4. Cost — Within budget?
  5. Strengths — Task-specific optimizations?

Output: Ranked list of providers with scores.


⚠️ ALERT LEVELS

Level Meaning Action
🟢 OK <60% Normal operation
🟡 WARNING 60-80% Monitor closely
🔴 CRITICAL 80-100% Consider fallback
⚫ EXHAUSTED 100% Use fallback immediately
⏸️ COOLDOWN Rate limited Wait for cooldown

🔄 INTEGRATION POINTS

From LLM Router

const { ResourceMonitor } = require('./llm-resource-monitor.js');
const monitor = new ResourceMonitor();
const best = monitor.getBestProvider('code', 50000);

From Gemini Key Pool

// After 429 error, mark cooldown
monitor.setCooldown('gemini-1', new Date(Date.now() + 60000));

Record After API Call

monitor.record('groq', 2500);

📁 DATA FILES

File Purpose
config/llm-resource-state.json Current usage state
config/llm-providers.json Provider configurations
config/llm-capability-scan.json Last capability scan

💡 BEST PRACTICES

For Maximum Capacity

  1. Create 4+ Google projects for Gemini — Each = 1,500 req/day
  2. Use MiniMax for coding — Unlimited, fast
  3. Use Groq for quick tasks — Fast, free
  4. Reserve Gemini for deep research — 1M context when needed

Model Selection Quick Guide

Heavy coding → Codex (GPT-5.4) or MiniMax
Fast tasks → Groq
Deep research → Gemini 3.1 Pro (1M context)
Reasoning/debugging → DeepSeek R1
Budget tasks → DeepSeek or OpenRouter

Related Skills

  • free-llm-providers — Provider details and API setup
  • subagent-launcher — Subagent orchestration
  • xavier-memory — Xavier memory integration

Version: 1.0.0 Unified: 2026-04-23 Status: ✅ OPERATIONAL

04 同仓库更多 Skills