arc-observing

You can detect and manage recurring AI agent behavior patterns by running a background observer daemon that captures tool calls, analyzes them for recurring sequences, and saves them as markdown instincts. You then confirm or contradict these instincts to influence future sessions.

AI instinct detection works by using a background daemon to capture tool calls and an LLM to analyze the captured sequences. When a behavioral pattern is identified, it is saved as a markdown file with a confidence score that you can manually confirm or contradict.

Yes, you can manage project-specific and global AI instincts separately. The skill supports both scopes, allowing you to apply behavioral patterns globally or restrict them to the current project's context.

You control the background daemon observing AI tool calls using specific commands to start, stop, and check its status. This daemon runs continuously to capture tool usage data for pattern analysis.

Unconfirmed AI behavioral instincts experience confidence decay over time. The skill includes a confidence decay mechanism that reduces the influence of older, unverified patterns, and an archiving system to manage them.

You reinforce or correct detected AI tool usage sequences by confirming or contradicting the instincts generated by the observer. Confirming a pattern increases its confidence score and auto-loading behavior, while contradicting it prevents future automatic application.