Multi-Provider Resilience

Lessons from hitting the Kimi K2.5 cap at 2 PM on a Wednesday.

The Crisis

Wednesday afternoon. Kimi K2.5 API calls start failing with rate limit errors. Primary provider capped out mid-conversation.

For an always-on agent, this is a full outage. No model = no responses = dead service.

The Fix (In Real-Time)

Step 1: Emergency Backup Plan

• Purchased second Kimi K2.5 plan
• Configured as kimi-backup provider in OpenClaw
• Set fallback chain: primary → backup → Claude → GPT

Step 2: The Proxy Problem

Kimi Code API requires specific client signatures ("coding harness" User-Agent). Direct API calls from OpenClaw were rejected.

Solution: Local proxy at loopback that:

• Receives OpenClaw requests
• Rewrites User-Agent to match Kimi Code requirements
• Forwards to Kimi Code endpoint
• Returns responses transparently

Step 3: Tool Execution Breakage

The proxy initially broke tool execution. Message role transformations (developer→system) were stripping tool_calls from assistant messages and converting tool roles incorrectly.

Fixed by narrowing the transformation scope:

• Only transform developer → system
• Leave tool_calls and tool role messages untouched
• Preserve conversation structure for the model

What Worked

Key Insight

Provider-specific restrictions are arbitrary. Kimi Code requires "coding harness" User-Agent. OpenAI requires specific message formats. Anthropic has its own quirks.

A proxy layer that normalizes these differences is essential for multi-provider setups.

Recommendations

1. Always Configure Fallbacks

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "kimi-coding/k2p5",
        "fallbacks": [
          "kimi-backup/k2p5",
          "anthropic/claude-opus-4-5",
          "openai/gpt-5.2"
        ]
      }
    }
  }
}

2. Keep Backup Credentials Ready

Don't scramble during an outage. Have backup API keys:

• In your password manager
• Synced to your VPS
• Ready to activate in seconds

3. Test Fallbacks Regularly

Force fallback by:

• Setting primary model to invalid key (test auth failure)
• Setting primary to non-existent model (test routing)
• Verifying each fallback tier works

4. Document Proxy Requirements

If your provider requires client spoofing:

• Document the exact headers/roles needed
• Version control your proxy config
• Test tool execution after any proxy changes

The Surprising Benefit

The backup plan didn't just prevent outage — it improved performance. The proxy's connection handling turned out to be more efficient than direct API calls. Latency dropped slightly after the switch.

Crisis → Opportunity → Better Architecture