openclaw - 💡(How to fix) Fix [Bug]: Agent-level model config ignored at session initialization (2026.5.7)

Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

## Summary

Agent-specific `model` configuration is ignored during session initialization. All sessions use `agents.defaults.model.primary` instead of the agent's configured model, making multi-agent setups with different models per agent impossible.

Or even more concise:

Agent-level model config ignored at session init - all agents use global default model

Pick whichever fits the form character limit! 🐈‍⬛

Steps to reproduce

## Steps to Reproduce

1. **Configure global default model:**
   ```json
   {
     "agents": {
       "defaults": {
         "model": {
           "primary": "anthropic/claude-sonnet-4-5"
         }
       }
     }
   }

1. Create an agent with a different model:

   {
     "agents": {
       "list": [
         {
           "id": "test-agent",
           "name": "test-agent",
           "workspace": "~/.openclaw/workspace-test",
           "model": "ollama/qwen2.5-coder:7b"
         }
       ]
     }
   }

2. Register the Ollama model in allowlist:

   {
     "agents": {
       "defaults": {
         "models": {
           "ollama/qwen2.5-coder:7b": {}
         }
       }
     }
   }

3. Ensure Ollama is running and model is available:

   curl -s http://localhost:11434/api/generate -X POST \
     -d '{"model":"qwen2.5-coder:7b","prompt":"test","stream":false}'

   Should return a valid response.

4. Bind the agent to a channel (e.g., Discord):

   {
     "bindings": [
       {
         "agentId": "test-agent",
         "match": {
           "channel": "discord",
           "peer": {
             "kind": "channel",
             "id": "YOUR_CHANNEL_ID"
           }
         }
       }
     ]
   }

5. Restart the Gateway:

   openclaw gateway restart

6. Send a message to the bound channel:

   what model are you running?

7. Check the session file:

   cat ~/.openclaw/agents/test-agent/sessions/sessions.json | grep -A 2 "modelOverride"

8. Check the trajectory file:

   head -5 ~/.openclaw/agents/test-agent/sessions/*.jsonl

Expected Result

• Session should use provider: "ollama", modelId: "qwen2.5-coder:7b"
• Agent responds: "I'm running Qwen 2.5 Coder 7B"

Actual Result

• Session shows:

  "modelOverride": "claude-sonnet-4-5",
  "providerOverride": "anthropic",
  "modelOverrideSource": "auto"

• Agent responds: "I'm running Claude Sonnet 4.5 (anthropic/claude-sonnet-4-5)"
• Trajectory shows model set to Anthropic immediately at session creation (no Ollama attempt)

Clean, reproducible steps that anyone can follow! 🐈‍⬛

### Expected behavior

```markdown
## Expected Behavior

When an agent is configured with a specific `model`:

```json
{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-sonnet-4-5"
      }
    },
    "list": [
      {
        "id": "test-agent",
        "model": "ollama/qwen2.5-coder:7b"
      }
    ]
  }
}

Sessions for that agent should:

1. Initialize with the agent's configured model (ollama/qwen2.5-coder:7b)
2. Have NO modelOverride / providerOverride fields (or they should be null)
3. Only set modelOverrideSource: "auto" if a runtime fallback occurs during an actual request

The agent-level model config should take precedence over agents.defaults.model.primary, per the documented model resolution cascade.

Trajectory should show:

• Ollama connection attempt at first request
• If Ollama succeeds: continue with Ollama
• If Ollama fails: fall back to configured fallback (if any), THEN set modelOverrideSource: "auto"

CLI should reflect:

$ openclaw status
Sessions:
│ agent:test-agent:discord:... │ group │ 1m ago │ ollama/qwen2.5-coder:7b │

Actual behavior

## Actual Behavior

Sessions are created with `agents.defaults.model.primary` regardless of agent-specific `model` config:

```json
{
  "modelOverride": "claude-sonnet-4-5",
  "providerOverride": "anthropic",
  "modelOverrideSource": "auto"
}

What happens:

1. Session is created for the agent
2. Model is immediately set to the global default (Anthropic) at session initialization
3. modelOverrideSource: "auto" is set without any runtime fallback occurring
4. Agent's configured model (ollama/qwen2.5-coder:7b) is never attempted
5. No Ollama connection attempt is visible in trajectory logs

Trajectory timeline:

T+0.000s: Session created
T+0.258s: Model set to "anthropic/claude-sonnet-4-5" (immediately)

No model_fallback event. No Ollama error. Just instant default override.

CLI shows wrong model:

$ openclaw status
Sessions:
│ agent:test-agent:discord:... │ group │ 1m ago │ claude-sonnet-4-5 │
                                                   ^^^^^^^^^^^^^^^^^^
                                                   Should be: ollama/qwen2.5-coder:7b

Agent confirms it's using wrong model:

I'm running Claude Sonnet 4.5 (anthropic/claude-sonnet-4-5).
The default model for this setup is ollama/qwen2.5-coder:7b, but this session is using Sonnet 4.5.

Result: Agent-level model config is loaded (visible in trajectory config snapshot) but completely ignored during session initialization.

OpenClaw version

2026.5.7

Operating system

macOS 12.7.4

Install method

Installation Method - Method: npm (global install via nvm) - Command: npm install -g openclaw - Node: v24.15.0 (managed by nvm) - Location: ~/.nvm/versions/node/v24.15.0/lib/node_modules/openclaw

Model

Model Intended: ollama/qwen2.5-coder:7b (configured for agent) Actual: anthropic/claude-sonnet-4-5 (what sessions used due to bug) Both models are relevant to this issue: - Agent was configured to use Ollama - Sessions were created with Anthropic (the bug)

Provider / routing chain

Provider / Routing Chain Discord → OpenClaw Gateway → Agent binding → Session initialization (where bug occurs) Specific routing: ```json { "bindings": [ { "agentId": "homie", "match": { "channel": "discord", "peer": { "kind": "channel", "id": "1502419400232407122" } } } ] }

Additional provider/model setup details

## Additional Provider/Model Setup Details

### Ollama Provider Configuration
```json
{
  "models": {
    "providers": {
      "ollama": {
        "baseUrl": "http://127.0.0.1:11434",
        "api": "ollama",
        "apiKey": "",
        "enabled": true
      }
    }
  }
}

Ollama status:

• Running locally on port 11434
• Models loaded with -1 keep-alive (never unload)
• Verified working via direct API test:

  curl -s http://localhost:11434/api/generate -X POST \
    -d '{"model":"qwen2.5-coder:7b","prompt":"test","stream":false}'
  # Returns valid response

Anthropic Provider Configuration

{
  "models": {
    "providers": {
      "anthropic": {
        "enabled": true
      }
    }
  }
}

Auth profile: anthropic:default (API key configured, working)

Model Registration (Allowlist)

{
  "agents": {
    "defaults": {
      "models": {
        "anthropic/claude-sonnet-4-5": {},
        "anthropic/claude-opus-4-7": {},
        "anthropic/claude-haiku-4": {},
        "ollama/qwen2.5-coder:7b": {},
        "ollama/qwen2.5:7b": {},
        "ollama/qwen2.5:3b": {}
      }
    }
  }
}

Agent Configuration Tested

Both string and object forms exhibit the same bug:

String form:

{
  "id": "homie",
  "model": "ollama/qwen2.5-coder:7b"
}

Object form:

{
  "id": "homie",
  "model": {
    "primary": "ollama/qwen2.5-coder:7b",
    "fallbacks": []
  }
}

Default Model Configuration

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-sonnet-4-5",
        "fallbacks": ["anthropic/claude-opus-4-7"]
      }
    }
  }
}

This is what all agents incorrectly fall back to, regardless of their individual config.

Complete technical context! 🐈‍⬛

### Logs, screenshots, and evidence

```shell

Impact and severity

## Impact

**High** - Core multi-agent functionality is blocked

**Affects:**
- All multi-agent setups attempting to use different models per agent
- Privacy-sensitive deployments requiring local-only models for specific agents
- Budget management strategies (local models for high-volume, cloud for quality)
- Any deployment relying on documented per-agent model configuration

**Cannot:**
- Build agent crews with specialized models (e.g., local for email/finance, cloud for creative)
- Enforce privacy boundaries (all agents fall back to cloud default)
- Test/implement documented agent model configuration features

**Workarounds:** None viable for production use
- Separate OpenClaw instances: Too heavy, complex management
- Global default rotation: Breaks other agents
- Manual `/model` overrides: Not persistent across sessions

**Severity:** High - Documented core feature completely non-functional, blocks entire use case with no practical workaround

Additional information

bug-report-agent-model-ignored.md