Config Change Order: Agent Model Fallback Chains

Date: 2026-04-20
Author: Daedalus 🎨
For: Socrates 🧠 to apply
Approved by: Matt

Architecture Principle

The Beelink runs Ollama as a cloud proxy only — no local inference. All fallback chains end with the Gaming PC (ollama-remote/) as the local lifeboat.

Changes

1. Wadsworth (main agent)

Current:

{
  "id": "main",
  "workspace": "~/.openclaw/workspace",
  "model": {
    "primary": "ollama/glm-5.1:cloud",
    "fallbacks": [
      "ollama/kimi-k2.5:cloud",
      "ollama-remote/phi4:14b"
    ]
  }
}

No change needed — already correct.

2. Daedalus

Current:

{
  "id": "daedalus",
  "workspace": "~/.openclaw/workspace-daedalus",
  "model": {
    "primary": "ollama/glm-5.1:cloud",
    "fallbacks": [
      "ollama/kimi-k2.5:cloud",
      "ollama-remote/phi4:14b"
    ]
  }
}

No change needed — already correct.

3. Socrates — FIX DUPLICATE FALLBACK

Current (broken):

{
  "id": "socrates",
  "workspace": "~/.openclaw/workspace-socrates",
  "model": {
    "primary": "ollama/kimi-k2.5:cloud",
    "fallbacks": [
      "ollama/kimi-k2.5:cloud",
      "ollama-remote/phi4:14b"
    ]
  }
}

After:

{
  "id": "socrates",
  "workspace": "~/.openclaw/workspace-socrates",
  "model": {
    "primary": "ollama/kimi-k2.5:cloud",
    "fallbacks": [
      "ollama/glm-5.1:cloud",
      "ollama-remote/phi4:14b"
    ]
  }
}

Summary Table

Active Memory

Per Matt — leave active-memory decisions to Socrates. Not included in this change order.

After Applying

1. Restart gateway: openclaw gateway restart
2. Verify all three agents respond on primary models
3. Monitor logs for any unexpected fallback events

Notes

• The qwen3.5:4b, llama3.2:3b, etc. shown in local ollama list are not valid fallback targets per architecture (Beelink = cloud proxy only)
• All three agents share the same Gaming PC lifeboat (phi4:14b) as last resort