divergence-monitor

Divergence Monitor Skill

When to Use

• User asks "why did I get a divergence alert?"
• User wants to check the current divergence verdict (CLEAN, DIVERGENCE, NO_MODEL, STALE)
• User asks about the behavioral model or what EDAMAME thinks Claude Desktop is doing
• User wants to force-refresh the behavioral model from recent transcripts
• User asks about the divergence engine status or history

Instructions

1. Start by calling edamame_claude_desktop_control_center to get the full status snapshot including:
   • Pairing and endpoint health
   • Divergence engine running state
   • Current behavioral model presence and last update time
   • Latest divergence verdict
2. For deeper diagnosis, use these EDAMAME MCP tools:
   • get_divergence_verdict for the current verdict with details
   • get_divergence_history for recent verdict timeline
   • get_behavioral_model for the current merged behavioral model
   • get_divergence_engine_status for engine configuration and health
3. If the verdict is DIVERGENCE, explain what the engine observed versus what the model predicted. Check which sessions or processes caused the mismatch.
4. If the verdict is NO_MODEL, guide the user to export intent:
   • Use edamame_claude_desktop_control_center_refresh_now to push current transcripts
   • Wait for the model to appear (EDAMAME processes it with its LLM provider)
5. If the verdict is STALE, the model has not been updated recently. Trigger a refresh.

Important

• Verdicts are authoritative from EDAMAME only. Never fabricate or cache verdicts locally.
• The behavioral model is generated server-side by EDAMAME's LLM provider from raw transcripts this package sends. The package itself does not run LLM inference.
• Multiple agent contributors (Claude Desktop, Cursor, OpenClaw, etc.) may be merged into one model. The Claude Desktop slice is identified by agent_type=claude_desktop.