Summary

The heartbeat.model config field is documented to allow per-heartbeat model overrides, but the setting is ignored at runtime. Heartbeats always use the main session's default model instead.

Environment

• OpenClaw version: 2026.2.9
• OS: Linux 6.8.0-100-generic
• Node: v22.22.0 / v25.6.0

Configuration

openclaw.json (relevant parts)

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "moonshot/kimi-k2.5"
      },
      "models": {
        "moonshot/kimi-k2.5": { "alias": "kimi" },
        "ollama/llama3.2:3b": { "alias": "local" }
      },
      "heartbeat": {
        "every": "1h",
        "model": "ollama/llama3.2:3b",
        "includeReasoning": false,
        "target": "last",
        "prompt": "Check: Any blockers, opportunities, or progress updates needed?"
      }
    }
  },
  "models": {
    "providers": {
      "ollama": {
        "baseUrl": "http://localhost:11434/v1",
        "api": "openai-completions",
        "models": [
          {
            "id": "llama3.2:3b",
            "name": "Llama 3.2 3B"
          }
        ]
      }
    }
  }
}

Expected Behavior

Heartbeats should run using ollama/llama3.2:3b (local, free) instead of moonshot/kimi-k2.5.

Actual Behavior

Heartbeats continue to use the main session's default model (kimi-k2.5).

Evidence

Config shows model is set

$ openclaw config get agents.defaults.heartbeat
{
  "every": "1h",
  "model": "ollama/llama3.2:3b",
  "includeReasoning": false,
  "target": "last",
  "prompt": "Check: Any blockers, opportunities, or progress updates needed?"
}

Logs show wrong model being used

2026-02-11T21:21:07.130Z debug agent/embedded embedded run start: 
  runId=ee265224-6ce0-41d1-bc39-6adcb78afb3b 
  sessionId=df3b02a4-600e-4f08-8c23-6ef2fe812307 
  provider=moonshot 
  model=kimi-k2.5 
  thinking=off 
  messageChannel=heartbeat

Note: messageChannel=heartbeat confirms this was a heartbeat run, yet it used kimi-k2.5 instead of the configured ollama/llama3.2:3b.

What was tried

1. Verified Ollama is running and model is available:

   $ curl http://localhost:11434/api/tags
   # llama3.2:3b is present and working

2. Added model alias to agents.defaults.models:

   "ollama/llama3.2:3b": { "alias": "local" }

3. Restarted gateway multiple times — config reloads successfully but model override is still ignored.

Workaround

Use cron jobs with --session isolated — model overrides work correctly in isolated sessions.

Impact

Users cannot use cheaper/local models for heartbeats as documented, forcing unnecessary API costs for simple periodic checks.

Documentation Reference

https://docs.openclaw.ai/gateway/heartbeat (states model field should override heartbeat model)