hermes - 💡(How to fix) Fix bug(honcho): background review agent writes system prompts as user messages, corrupting peer representations

Bug

The background self-improvement review agent (_spawn_background_review() in run_agent.py) initializes external memory providers (e.g. Honcho) because skip_memory defaults to False. The review prompt is then written as a user message to the external provider, polluting peer representations with agent-internal system instructions.

Root cause

1. _spawn_background_review() creates a new AIAgent without skip_memory=True
2. The Honcho plugin's cron guard checks agent_context in ("cron", "flush") or platform == "cron" — but the review agent inherits platform="discord" from the parent and agent_context defaults to "primary"
3. The plugin activates, and sync_turn() writes the self-improvement prompt as a user message
4. With pinPeerName: true, this is attributed to the human peer, corrupting their representation

Impact

• User representations accumulate false observations about skill management preferences
• Cross-session context becomes unreliable
• The Honcho deriver processes agent-internal instructions as if they were user statements

Expected behavior

The background review agent should not write to external memory providers. It only needs the built-in memory and skills tools.

Fix

Pass skip_memory=True to the review agent constructor. Also extend the Honcho cron guard to cover "background_review" and "subagent" contexts (matching supermemory plugin behavior).

PR incoming.