360 Secure SkillHub
360 安全技能中心

Cron Model Fix

针对 OpenClaw cron 任务中模型覆盖被拒绝、回退到 agent 默认模型的问题,提供日志排查、配置层检查与修复步骤:把缺失模型加入 agents.defaults.models 允许列表,重启 gateway 并验证运行结果;附带配置说明文档与 Python 修复脚本。

mklue
mklue
热度 38.10
最近更新 2026-05-11 17:38:22
分类:开发依赖:代码执行 / 读写文件应用场景:调试排错 / DevOps / 工作流

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add cron-model-fix
Metadata
name
cron-model-fix
description
Diagnose and fix OpenClaw cron job model override issues. Use when cron jobs show "not allowed, falling back to agent defaults" in logs, experience unexpected cloud token burn, have slow run times, or models aren't being applied correctly. Fixes agent model allowlist configuration.

Cron Model Fix

Diagnose and fix OpenClaw cron job model override issues where configured models are rejected and fall back to agent defaults.

When to Use This Skill

Use when:

  • Cron jobs show "not allowed, falling back to agent defaults" in gateway logs
  • Unexpected cloud token burn on cron jobs configured with local models
  • Slow cron run times (60-180s) when local models should be used
  • Model overrides in cron jobs aren't being applied
  • Need to validate agent model allowlist configuration

Quick Start

# Diagnose the issue
openclaw skill run cron-model-fix --diagnose

# Apply the fix (adds missing models to allowlist)
openclaw skill run cron-model-fix --fix --model ollama/qwen3.5:0.8b

# Validate configuration
openclaw skill run cron-model-fix --validate

Root Cause

OpenClaw has three model configuration layers. ALL must include the model:

Layer 1: Provider Model Definition

Location: models.providers.<provider>.models[] Purpose: Defines model specs (context, costs, capabilities)

Layer 2: Agent Default Model

Location: agents.defaults.model.primary Purpose: Default model when none specified

Layer 3: Agent Model Allowlist ← COMMON ISSUE

Location: agents.defaults.models Purpose: WHICH MODELS ARE PERMITTED for agent/cron use

Problem: Model exists in Layer 1, but missing from Layer 3 (allowlist).

Gateway logs show:

{"subsystem":"cron"}
"payload.model 'ollama/qwen3.5:0.8b' not allowed, falling back to agent defaults"

Diagnosis

Step 1: Check Gateway Logs

tail -100 /tmp/openclaw/openclaw-*.log | grep -i "not allowed\|falling back"

If you see: "payload.model '<model>' not allowed, falling back to agent defaults" → Model is missing from agent allowlist

Step 2: Check Agent Allowlist

cat ~/.openclaw/openclaw.json | python3 -c "
import json, sys
config = json.load(sys.stdin)
models = config.get('agents', {}).get('defaults', {}).get('models', {})
print('Allowed models:')
for model in models:
    print(f'  - {model}')
"

Check if your cron model is in the list. If not, it will be rejected.

Step 3: Check Cron Job Configuration

openclaw cron list

Verify:

  • Cron job has model specified in payload
  • Model format matches allowlist (e.g., ollama/qwen3.5:0.8b)

The Fix

Add the model to agents.defaults.models in ~/.openclaw/openclaw.json:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "ollama/qwen3.5:cloud",
        "fallbacks": [...]
      },
      "models": {
        "ollama/glm-4.7-flash": {},
        "ollama/kimi-k2.5:cloud": {},
        "ollama/qwen2.5vl:7b": {},
        "ollama/qwen3.5:397b-cloud": {},
        "ollama/qwen3.5:4b-32K": {},
        "ollama/qwen3.5:4b-32k": {},
        "ollama/qwen3.5:9b-128k": {},
        "ollama/qwen3.5:cloud": {},
        "ollama/qwen3.5:0.8b": {},  ← ADD THIS
        "ollama/qwen3.5:2b": {}     ← Optional
      }
    }
  }
}

Manual Fix Steps

  1. Edit config:

    nano ~/.openclaw/openclaw.json
    
  2. Add to agents.defaults.models:

    "ollama/qwen3.5:0.8b": {},
    
  3. Restart gateway:

    openclaw gateway restart
    
  4. Verify:

    openclaw cron runs --id <job-id> --limit 3
    

    Look for "model": "qwen3.5:0.8b" instead of qwen3.5:cloud

Automated Fix Script

Use the included script:

python3 ~/.npm-global/lib/node_modules/openclaw/skills/cron-model-fix/scripts/add-model-allowlist.py --model ollama/qwen3.5:0.8b

Expected Results

Before Fix

Metric Value
Model Cloud fallback
Duration 60-180 seconds
Input Tokens 200K-600K per run
Cost Cloud token burn

After Fix

Metric Value Improvement
Model Local (e.g., qwen3.5:0.8b) ✅ Free
Duration 1-13 seconds 5-14x faster
Input Tokens 5K-36K 85-95% reduction
Cost ZERO 100% savings

Validation

After applying fix, verify:

# Check for model rejection warnings
tail -50 /tmp/openclaw/openclaw-*.log | grep -i "not allowed"

# Should be EMPTY (no warnings)
# Check cron run history
openclaw cron runs --id <job-id> --limit 1

# Should show:
# "model": "qwen3.5:0.8b" (not cloud)
# "durationMs": <15000 (fast)
# "input_tokens": <50000 (low)

Common Issues

Issue: Model still not working after fix

Check:

  1. Gateway restarted after config change?
  2. Model format matches exactly? (e.g., ollama/qwen3.5:0.8b vs qwen3.5:0.8b)
  3. Model exists in provider config (Layer 1)?

Issue: Invalid JSON after editing

Fix:

openclaw doctor --fix

Or restore from backup:

cp ~/.openclaw/openclaw.json.backup ~/.openclaw/openclaw.json

Related Skills

  • inbox-optimizer - Optimizes inbox scanning patterns (mesh-specific)
  • healthcheck - General OpenClaw system health monitoring

References

  • See references/model-config-layers.md for detailed configuration structure
  • See references/troubleshooting-examples.md for real-world case studies

Version

1.0.0 - Initial release (2026-03-30)