Skill #85: Model Failover Manager

Enterprise-grade AI provider rotation, cooldowns, and automatic failover for K.I.T.'s trading decisions.

Why Model Failover?

AI providers have rate limits, outages, and billing issues. K.I.T. needs:

• 24/7 uptime for autonomous trading
• Automatic recovery from provider failures
• Cost optimization across providers
• Quality maintenance when switching models

Features

Multi-Provider Support

Provider	Models	Rate Limit Handling
Anthropic	Claude Opus, Sonnet, Haiku	Per-minute, per-day
OpenAI	GPT-4o, GPT-4-turbo, o1	TPM, RPM
Google	Gemini 2.0, 1.5 Pro	Per-minute
xAI	Grok 2	Per-minute
DeepSeek	DeepSeek V3, R1	Per-minute
Groq	Llama, Mixtral	Per-minute, free tier
OpenRouter	All models	Aggregated
Local	Ollama, vLLM	No limits

Failover Strategies

1. Round-Robin Rotation Distributes load across providers:

Request 1 → Anthropic
Request 2 → OpenAI
Request 3 → Google
Request 4 → Anthropic (back to start)

2. Priority Cascade Falls back through priority order:

Primary:   Claude Opus 4
Fallback1: GPT-4o
Fallback2: Gemini 2.0
Fallback3: Local Ollama (never fails)

3. Cost-Optimized Routes to cheapest available provider:

Simple queries → Haiku ($0.25/1M)
Complex analysis → Sonnet ($3/1M)
Critical decisions → Opus ($15/1M)

4. Latency-Optimized Tracks response times and routes to fastest:

Groq:     ~200ms (when available)
Anthropic: ~800ms
OpenAI:   ~1200ms

Cooldown System

Exponential backoff for failures:

1st failure:  1 minute cooldown
2nd failure:  5 minutes
3rd failure:  25 minutes
4th+ failure: 1 hour (cap)

Separate handling for:

• Rate limits: Standard cooldown + retry
• Auth errors: Immediate failover, longer cooldown
• Billing issues: 5-hour initial backoff, doubles each time
• Timeouts: Shorter cooldown (30 seconds)

Session Stickiness

K.I.T. pins a provider per trading session to:

• Keep provider caches warm
• Maintain conversation context
• Avoid inconsistent analysis

Re-pins only when:

• Session resets
• Current provider enters cooldown
• User manually switches

Usage

kit model status              # Show provider health
kit model failover --test     # Simulate failover
kit model cooldown anthropic  # Manually cooldown a provider
kit model priority            # Show/edit fallback order

Configuration

{
  "ai": {
    "defaultProvider": "anthropic",
    "defaultModel": "claude-opus-4-5-20251101",
    "failover": {
      "enabled": true,
      "strategy": "priority",
      "fallbacks": [
        "anthropic/claude-opus-4-5-20251101",
        "openai/gpt-4o",
        "google/gemini-2.0-flash-exp",
        "ollama/llama3.3"
      ],
      "cooldowns": {
        "rateLimit": [60, 300, 1500, 3600],
        "authError": 3600,
        "billing": 18000,
        "timeout": 30
      },
      "sessionSticky": true,
      "costOptimize": false
    },
    "providers": {
      "anthropic": { "apiKey": "sk-ant-..." },
      "openai": { "apiKey": "sk-..." },
      "google": { "apiKey": "AIza..." }
    }
  }
}

Health Dashboard

🧠 AI Provider Status

Provider      Model              Status    Latency   Cooldown
────────────────────────────────────────────────────────────
anthropic     claude-opus-4.5    ✅ OK     823ms     -
openai        gpt-4o             ✅ OK     1.2s      -
google        gemini-2.0         ⚠️ COOL   -         3m left
xai           grok-2             ✅ OK     650ms     -
deepseek      deepseek-v3        ❌ BILLING -        4h left
ollama        llama3.3           ✅ LOCAL  180ms     -

Session: Pinned to anthropic (42 requests)
Strategy: Priority cascade

Trading Integration

High-Stakes Decisions

For critical trading decisions, force best model:

const decision = await kit.ai.complete({
  prompt: "Should I exit this BTC position?",
  model: "claude-opus-4-5-20251101",
  failover: false  // Don't downgrade for critical decisions
});

Parallel Consensus

For important decisions, query multiple models:

const opinions = await kit.ai.consensus({
  prompt: "Analyze NVDA earnings impact",
  models: ["claude-opus-4-5", "gpt-4o", "gemini-2.0"],
  threshold: 2/3  // Need 2 of 3 to agree
});

Cost-Aware Routing

// Route based on task complexity
const response = await kit.ai.smart({
  prompt: userMessage,
  complexity: 'auto',  // Analyzes prompt to choose model
  maxCost: 0.01       // Cap at $0.01 per request
});

Metrics & Logging

Tracks per provider:

• Request count / success rate
• Average latency (p50, p95, p99)
• Token usage / cost
• Error types and frequencies
• Cooldown history

Daily report:

📊 AI Provider Report (24h)

Provider     Requests  Success  Latency  Cost
─────────────────────────────────────────────
anthropic    847       99.2%    812ms    $2.45
openai       156       98.7%    1.1s     $0.87
google       23        91.3%    650ms    $0.12
local        412       100%     180ms    $0.00

Total: 1438 requests, $3.44 cost
Failovers: 12 (0.8%)

Related Skills

• #11 LLM Agent Core: Base AI integration
• #52 AI Screener: Uses AI for stock screening
• #57 Economic Calendar: AI interprets events
• #83 Deal Manager: AI-powered position management

References

• OpenClaw Model Failover: /concepts/model-failover.md
• Auth Profile Rotation: /concepts/oauth.md
• Provider Docs: /providers/