Heartbeat mechanism issues: model switching, trigger failures, and output format changes

Problem Description

I'm experiencing several issues with the OpenClaw heartbeat mechanism:

1. Inconsistent triggering - Configured for 120-minute intervals, but heartbeat doesn't always trigger on schedule
2. Model switching - Configured to use DeepSeek model, but sometimes executes with Kimi instead
3. Output format changes - Previously heartbeat output showed "thinking process first, then execution process", but now only shows "execution process" (tool outputs + final HEARTBEAT_OK)

Configuration

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "moonshot/kimi-k2.5",
        "fallbacks": [
          "qwen-portal/coder-model",
          "qwen-portal/vision-model",
          "openrouter/qwen/qwen3-max",
          "moonshot/kimi-k2.5",
          "deepseek/deepseek-chat"
        ]
      },
      "heartbeat": {
        "every": "120m",
        "activeHours": {
          "start": "08:00",
          "end": "02:00"
        },
        "model": "deepseek/deepseek-chat"
      }
    }
  }
}

Expected Behavior

1. Heartbeat should trigger consistently at the configured 120-minute interval
2. Heartbeat should always use the configured deepseek/deepseek-chat model
3. Heartbeat output should include both thinking process (analysis after reading HEARTBEAT.md) and execution process

Actual Behavior

1. Heartbeat triggering is unstable - sometimes doesn't trigger after 120 minutes
2. Observed heartbeat executions sometimes use moonshot/kimi-k2.5 instead of configured DeepSeek
3. Output format changed to only show tool calls and final HEARTBEAT_OK, missing the thinking/analysis phase

Environment

• OpenClaw version: 2026.3.24 (from meta.lastTouchedVersion)
• OS: Linux (WSL2)
• Channel: Feishu

Additional Context

I noticed that agents.defaults.model.primary is set to moonshot/kimi-k2.5, while heartbeat.model is set to deepseek/deepseek-chat. Could the heartbeat execution be affected by the primary model configuration?

Also, the output format change appeared recently - previously it worked correctly with two-phase output (thinking + execution).

Thanks for the great work on OpenClaw! It's an awesome project. Let me know if you need any logs or additional testing.