System prompts (cron, task reminders) can block agent from responding to user

Description

When using Claude Code as a long-running agent with cron-based polling (via CronCreate), system-injected prompts can consume context and block the agent from responding to actual user messages.

Steps to reproduce

1. Set up a recurring cron job (e.g., */2 * * * *) that checks for messages
2. Each cron tick injects a full prompt into the conversation
3. Even when results are empty (no messages to process), each tick adds ~500+ tokens of context
4. System reminders like "task tools haven't been used recently" also inject between turns
5. When the user sends a message, it can get buried between system prompts and the agent processes cron ticks instead of responding

Expected behavior

• System prompts should not block user messages
• Empty poll results should be ultra-lightweight or suppressed
• User messages should have priority over recurring system prompts

Actual behavior

• Cron prompts fire continuously and consume context even with empty results
• The agent ends up responding to system prompts instead of user messages
• Context window fills up faster, leading to compaction and lost conversation history

Environment

• Claude Code CLI, running as a long-running agent via tmux
• Using CronCreate for recurring tasks
• Communication via iMessage MCP server

Suggested improvements

1. Allow cron jobs to specify a "silent on empty" mode
2. Prioritize user-originated messages over system/cron prompts
3. Reduce context footprint of system reminders