openclaw - 💡(How to fix) Fix [Bug]: openclaw agent times out/hangs with local Ollama on macOS while TUI interactive chat works [1 participants]

  openclaw agent --agent main --message "Reply with exactly: MAC MINI CLAW OK"

  Task Kind Status Summary                                                                                           
  ... cli timed_out Reply with exactly: MAC MINI CLAW OK                                

### Expected behavior

Task to be executed. No timeout.  

### Actual behavior

Timout. Session is stuck. Wedged. 

### OpenClaw version

2026.4.7 (5050017)

### Operating system

macOS 26.3.1 arm64

### Install method

bash script from web site

### Model

qwen3.5:9b 6488c96fa5fa 6.6 GB

### Provider / routing chain

open claw -> Ollama local

### Additional provider/model setup details

- OpenClaw: 2026.4.7 (5050017)                                                                                       
- OS: macOS 26.3.1 arm64                                                                                             
- Node: 22.22.2                                                                                                      
- Ollama: 0.20.3                                                                                                     
- Host: Mac mini M4                                                                                                  
- Gateway mode: local                                                                                                
- Gateway service: launch agent                                                                                      
- Local model tested:                                                                                                
    - ollama/qwen3.5:9b                                                                                              
- Also tested:                                                                                                       
    - ollama/gemma4:e4b                                                                                              
                                                                                                                     
Local model availability                                                                                             
                                                                                                                     
Ollama is healthy and local models are present:                                                                      
                                                                                                                     
```bash                                                                                                              
  ollama list                                                                                                        
  NAME ID SIZE MODIFIED                                                                                              
  gemma4:26b 5571076f3d70 17 GB 35 hours ago                                                                         
  gemma4:e4b c6eb396dbd59 9.6 GB 35 hours ago                                                                        
  qwen3.5:9b 6488c96fa5fa 6.6 GB 2 days ago

  ollama run gemma4:e4b "Reply with exactly: MAC MINI CLAW OK"

  MAC MINI CLAW OK                                                           

### Logs, screenshots, and evidence

```shell
Summary                                                                                                              
                                                                                                                     
On macOS, with OpenClaw configured to use a local Ollama model, the interactive openclaw tui path works, but         
one-shot CLI agent runs via openclaw agent --agent main --message ... repeatedly time out or hang.                   
                                                                                                                     
This appears to affect the openclaw agent / gateway-injected / task path specifically, not the local Ollama model    
itself.                                                                                                              
                                                                                                                     
Environment                                                                                                          
                                                                                                                     
- OpenClaw: 2026.4.7 (5050017)                                                                                       
- OS: macOS 26.3.1 arm64                                                                                             
- Node: 22.22.2                                                                                                      
- Ollama: 0.20.3                                                                                                     
- Host: Mac mini M4                                                                                                  
- Gateway mode: local                                                                                                
- Gateway service: launch agent                                                                                      
- Local model tested:                                                                                                
    - ollama/qwen3.5:9b                                                                                              
- Also tested:                                                                                                       
    - ollama/gemma4:e4b                                                                                              
                                                                                                                     
Local model availability                                                                                             
                                                                                                                     
Ollama is healthy and local models are present:                                                                      
                                                                                                                     
                                                                                                              
  ollama list                                                                                                        
  NAME ID SIZE MODIFIED                                                                                              
  gemma4:26b 5571076f3d70 17 GB 35 hours ago                                                                         
  gemma4:e4b c6eb396dbd59 9.6 GB 35 hours ago                                                                        
  qwen3.5:9b 6488c96fa5fa 6.6 GB 2 days ago                                                                          
                                                                                                                  
                                                                                                                     
Direct Ollama usage works, for example:                                                                              
                                                                                                                     
                                                                                                              
  ollama run gemma4:e4b "Reply with exactly: MAC MINI CLAW OK"                                                       
                                                                                                                  
                                                                                                                     
returns:                                                                                                             
                                                                                                                     
                                                                                                              
  MAC MINI CLAW OK                                                                                                   
                                                                                                                  
                                                                                                                     
OpenClaw config used                                                                                                 
                                                                                                                     
Relevant config:                                                                                                     
                                                                                                                     
                                                                                                              
  {                                                                                                                  
    "gateway": {                                                                                                     
      "mode": "local",                                                                                               
      "auth": {                                                                                                      
        "mode": "token",                                                                                             
        "token": "..."                                                                                               
      }                                                                                                              
    },                                                                                                               
    "agents": {                                                                                                      
      "defaults": {                                                                                                  
        "model": {                                                                                                   
          "primary": "ollama/qwen3.5:9b",                                                                            
          "fallbacks": []                                                                                            
        }                                                                                                            
      }                                                                                                              
    },                                                                                                               
    "plugins": {                                                                                                     
      "entries": {                                                                                                   
        "ollama": { "enabled": true }                                                                                
      }                                                                                                              
    }                                                                                                                
  }                                                                                                                  
                                                                                                                  
                                                                                                                     
Also confirmed auth profile exists for Ollama:                                                                       
                                                                                                                     
                                                                                                              
  {                                                                                                                  
    "version": 1,                                                                                                    
    "profiles": {                                                                                                    
      "ollama:default": {                                                                                            
        "type": "api_key",                                                                                           
        "provider": "ollama",                                                                                        
        "key": "ollama-local"                                                                                        
      }                                                                                                              
    }                                                                                                                
  }                                                                                                                  
                                                                                                                  
                                                                                                                     
What works                                                                                                           
                                                                                                                     
### Interactive TUI works                                                                                            
                                                                                                                     
Running:                                                                                                             
                                                                                                                     
                                                                                                              
  openclaw tui                                                                                                       
                                                                                                                  
                                                                                                                     
and sending a simple prompt like:                                                                                    
                                                                                                                     
                                                                                                              
  hello                                                                                                              
                                                                                                                  
                                                                                                                     
gets a response successfully.                                                                                        
                                                                                                                     
This suggests:                                                                                                       
- gateway is healthy                                                                                                 
- local session path is healthy                                                                                      
- Ollama provider path is healthy                                                                                    
- local interactive chat is usable                                                                                   
                                                                                                                     
What fails                                                                                                           
                                                                                                                     
### One-shot CLI path times out or hangs                                                                             
                                                                                                                     
Running:                                                                                                             
                                                                                                                     
                                                                                                              
  openclaw agent --agent main --message "Reply with exactly: MAC MINI CLAW OK"                                       
                                                                                                                  
                                                                                                                     
creates a cli task but does not return successfully.                                                                 
                                                                                                                     
Observed outcomes:                                                                                                   
- tasks end as timed_out                                                                                             
- sometimes the process appears to hang                                                                              
- increasing timeout makes it hang longer, but still not complete                                                    
                                                                                                                     
Task table example:                                                                                                  
                                                                                                                     
                                                                                                              
  Task Kind Status Summary                                                                                           
  ... cli timed_out Reply with exactly: MAC MINI CLAW OK                                                             
                                                                                                                  
                                                                                                                     
Longer timeout test                                                                                                  
                                                                                                                     
Also tested:                                                                                                         
                                                                                                                     
                                                                                                              
  openclaw agent --agent main --timeout 120 --message "Reply with exactly: MAC MINI CLAW OK"                         
                                                                                                                  
                                                                                                                     
This changed behavior from “times out quickly” to “remains running longer”, but still did not complete successfully. 
                                                                                                                     
So --timeout affects behavior, but does not resolve the underlying issue.                                            
                                                                                                                     
Relevant logs                                                                                                        
                                                                                                                     
Repeated log lines:                                                                                                  
                                                                                                                     
                                                                                                              
  [agent] Profile ollama:default timed out. Trying next account...                                                   
  [agent] embedded run failover decision: runId=... stage=assistant decision=surface_error reason=timeout            
provider=openclaw/gateway-injected profile=...                                                                       
                                                                                                                  
                                                                                                                     
Later, after removing fallbacks:                                                                                     
                                                                                                                     
                                                                                                              
  [agent] embedded run failover decision: runId=... stage=assistant decision=surface_error reason=timeout            
provider=ollama/qwen3.5:9b profile=...                                                                               
                                                                                                                  
                                                                                                                     
Earlier, before fallbacks were disabled, the timeout path also cascaded into misleading OpenAI auth failures:        
                                                                                                                     
                                                                                                              
  No API key found for provider "openai"                                                                             
                                                                                                                  
                                                                                                                     
Setting:                                                                                                             
                                                                                                                     
                                                                                                              
  "fallbacks": []                                                                                                    
                                                                                                                  
                                                                                                                     
removed some of that noise, but did not fix the timeout/hang.                                                        
                                                                                                                     
Why this seems like an OpenClaw bug                                                                                  
                                                                                                                     
- Direct Ollama execution works                                                                                      
- TUI interactive session works                                                                                      
- The failure appears isolated to openclaw agent / one-shot CLI / task path                                          
- The timeout/failover behavior seems to happen in the OpenClaw orchestration layer rather than in Ollama itself     
                                                                                                                     
Expected behavior                                                                                                    
                                                                                                                     
openclaw agent --agent main --message "Reply with exactly: MAC MINI CLAW OK" should return a response successfully   
when the configured local Ollama model is healthy and interactive TUI chat is already working.                       
                                                                                                                     
Actual behavior                                                                                                      
                                                                                                                     
- CLI task times out or hangs                                                                                        
- logs show Ollama profile timeout / surface timeout behavior                                                        
- TUI remains usable, making this look path-specific rather than a full provider failure                             
                                                                                                                     
Additional notes                                                                                                     
                                                                                                                     
Local docs appear relevant here:                                                                                     
- openclaw agent supports --timeout <seconds>                                                                        
- docs also distinguish:                                                                                             
    - a shorter wait behavior (agent.wait)                                                                           
    - longer embedded runtime timeout (agents.defaults.timeoutSeconds )                                              
                                                                                                                     
This may point to a mismatch between the CLI wait/surface timeout and the actual local Ollama response path.