Bug Report: Ollama Models Return Broken Responses in OpenClaw

Summary

Ollama integration appears to connect successfully but returns corrupted/nonsensical responses instead of proper model output.

Environment

• OpenClaw Version: 2026.2.1 (ed4529e) → 2026.2.2-3 (updated during troubleshooting)
• OS: macOS (Darwin 25.2.0 arm64)
• Node: v25.5.0
• Ollama Version: Latest (service running on PID 1024)
• Models Tested: llama3.1:8b, mistral:7b

Expected Behavior

When using ollama models, should receive proper responses like "2+2 = 4"

Actual Behavior

• Model switching appears successful (session_status shows correct ollama model)
• Token counting works (shows proper token usage)
• But responses are completely broken/nonsensical

Example:

Input: "What is 2+2?"
Expected: "4" or "2+2 = 4"
Actual: Random text about sessions_send functions and memory_get operations

Troubleshooting Attempted

✅ Working Components:

• Ollama service running correctly (PID 1024)
• Direct API calls work: curl http://127.0.0.1:11434/api/generate returns proper responses
• Models available: llama3.1:8b, mistral:7b, qwen2.5:14b, llava:7b, codellama:7b
• Environment variable set: OLLAMA_API_KEY='ollama-local' in ~/.zshrc

✅ Configuration Fixes Applied:

• Added missing ollama:default auth profile to config:

"auth": {
  "profiles": {
    "ollama:default": {
      "provider": "ollama", 
      "mode": "api_key"
    }
  }
}

✅ Status Shows Connection Works:

🧠 Model: ollama/llama3.1:8b · 🔑 api-key ollama…-local (ollama:default)
🧮 Tokens: 12 in / 433 out

❌ But Responses Are Garbage:

• Model switching succeeds visually
• Token counting works
• Authentication appears successful
• Response content is completely broken

Reproduction Steps

1. Configure ollama models in OpenClaw config
2. Set OLLAMA_API_KEY='ollama-local'
3. Add ollama:default auth profile
4. Switch to any ollama model: /model ollama/llama3.1:8b
5. Ask simple question: "What is 2+2?"
6. Observe broken response instead of "4"

Diagnostic Info

$ openclaw status --all
🦞 OpenClaw 2026.2.1 (ed4529e)
🕒 Time: Wednesday, February 4th, 2026 — 6:08 AM (America/Phoenix)
🧠 Model: ollama/llama3.1:8b · 🔑 api-key ollama…-local (ollama:default)
🧮 Tokens: 11 in / 171 out
📚 Context: 25k/200k (13%) · 🧹 Compactions: 1
🧵 Session: agent:main:main • updated just now
⚙️ Runtime: direct · Think: off · elevated
🪢 Queue: collect (depth 0)

Hypothesis

This appears to be a bug in OpenClaw's ollama provider response parsing/handling, not a connection or authentication issue. The provider successfully connects to Ollama and counts tokens correctly, but the response processing pipeline is corrupted.

Impact

• Ollama models completely unusable in OpenClaw
• Users forced to use expensive API models instead of free local models
• Significant cost impact for high-usage scenarios

Workaround

Changed default model from ollama/llama3.1:8b to anthropic/claude-sonnet-4-20250514 to restore functionality.

Additional Context

Direct Ollama API testing confirms the service works perfectly:

curl -X POST http://127.0.0.1:11434/api/generate \
  -d '{"model": "llama3.1:8b", "prompt": "What is 2+2?", "stream": false}'
# Returns proper JSON with "4" response

The issue appears to be isolated to OpenClaw's ollama integration layer.