openclaw - 💡(How to fix) Fix Feature: Per-session tool scoping for cron jobs [1 comments, 2 participants]

Problem

Cron jobs running as isolated sessions inherit the agent's full tool configuration. For agents with many tools (e.g. 30+ Feishu tools), this means the system prompt + tool definitions can exceed 130K tokens, making it impossible to use smaller/cheaper models (e.g. glm-4.5-air with 131K context) for simple recurring tasks.

Real-world example: a project progress check cron task that only needs 3 tools (message, feishu_bitable_app_table_record, sessions_send) but inherits 30+ tool definitions, pushing total context to ~155K tokens and causing overflow on models with 131K context windows.

Proposed Solution

Allow cron job payload to specify a tool subset:

{
  "payload": {
    "kind": "agentTurn",
    "message": "Check project progress...",
    "tools": ["message", "feishu_bitable_app_table_record", "sessions_send"],
    "model": "zai/glm-4.5-air"
  }
}

When tools is specified, only those tool definitions are injected into the isolated session. When omitted, behavior is unchanged (all agent tools loaded).

Benefits

1. Cost optimization: Use smaller/cheaper models for simple cron tasks
2. Reduced latency: Smaller prompts = faster inference
3. Better context utilization: More room for actual task content instead of unused tool schemas
4. Security principle: Least privilege — cron tasks only get tools they need

Use Cases

• Progress check tasks (3-5 tools) running every 30 min
• Health check scripts (1-2 tools)
• Daily reports (5-10 tools)
• Any simple recurring task that doesn't need the full agent toolkit