Multi-model automatic fallback system for AI agents

Overview

This skill provides automatic model fallback functionality for OpenClaw agents. When the primary model fails (unavailable, slow, or rate-limited), it automatically switches to backup models in a predefined priority order.

Features

• Automatic Fallback: Seamlessly switch to backup models on failure
• Configurable Priority: Define your own model fallback order
• Health Monitoring: Track model availability and response times
• Cost Optimization: Use cheaper models for simple tasks
• Logging: Full audit trail of fallback events

Supported Models

Configuration

Default Fallback Chain

{
  "fallback_chain": [
    {
      "provider": "minimax-portal",
      "model": "MiniMax-M2.5",
      "priority": 1,
      "timeout": 30,
      "max_retries": 3
    },
    {
      "provider": "moonshot",
      "model": "kimi-k2.5",
      "priority": 2,
      "timeout": 30,
      "max_retries": 2
    },
    {
      "provider": "zhipu",
      "model": "glm-4-air",
      "priority": 3,
      "timeout": 20,
      "max_retries": 2
    }
  ]
}

Environment Variables

Usage

Basic Usage

The skill automatically handles model failures. No explicit calls needed.

# Trigger a model call (fallback happens automatically on failure)

Manual Fallback

# Force fallback to next model
/scripts/model-fallback.sh --force-next
# Check current model status
/scripts/model-fallback.sh --status
# Reset to primary model
/scripts/model-fallback.sh --reset

Configuration

Edit config.json to customize the fallback chain:

{
  "fallback_chain": [
    {"provider": "...", "model": "...", "priority": 1}
  ],
  "health_check": {
    "enabled": true,
    "interval_seconds": 300
  }
}

How It Works

1. User makes request with primary model
Model call fails (error, timeout, rate limit)
Skill detects failure
Wait 3 seconds (debounce)
Switch to next model in chain
Retry request with new model
If successful, return result
If failed, repeat steps 4-7
If all models fail, return error with details

Fallback Triggers

Logging

Logs are written to:

• ~/.openclaw/logs/model-fallback.log

Log Format

[2026-02-27 14:00:00] [INFO] Primary model MiniMax-M2.5 called
[2026-02-27 14:00:05] [WARN] Model failed: rate limit exceeded
[2026-02-27 14:00:05] [INFO] Falling back to Kimi K2.5
[2026-02-27 14:00:10] [INFO] Fallback successful

Cost Optimization

Use cheaper models for simple tasks:

{
  "task_routing": {
    "simple_query": ["glm-4-air", "glm-4-flash"],
    "complex_reasoning": ["MiniMax-M2.5", "kimi-k2.5"],
    "long_context": ["kimi-k2.5", "MiniMax-M2.1"]
  }
}

Integration

OpenClaw Configuration

Add to openclaw.json:

{
  "models": {
    "mode": "merge",
    "fallback": {
      "enabled": true,
      "config": "~/.openclaw/skills/model-fallback/config.json"
    }
  }
}

Health Check

Integrate with system health monitoring:

# Check model health
curl http://localhost:18789/api/models/health

Troubleshooting

Fallback Not Working

• Check if fallback is enabled: echo $MODEL_FALLBACK_ENABLED
• Verify config exists: ls ~/.openclaw/skills/model-fallback/config.json
• Check logs: tail -f ~/.openclaw/logs/model-fallback.log

Models Always Failing

• Check API keys are valid
• Verify network connectivity
• Check rate limits on provider dashboard

Examples

Example 1: Simple Fallback

User: "Hello"
System: Using MiniMax-M2.5...
System: Rate limited, switching to Kimi K2.5...
System: Response from Kimi K2.5: "Hello! How can I help?"

Example 2: Cost Optimization

User: "What is 2+2?"
System: Routing to glm-4-air (low cost)...
System: Response: "2+2=4"

Example 3: Long Document

User: "Summarize this 100-page PDF"
System: Detected long context requirement
System: Routing to Kimi K2.5 (256K context)...
System: Processing...

License

MIT

Author

CC (AI Assistant)

Version

1.0.0