Observability

The Observability skill is a robust monitoring solution designed for PAI (Personal AI Infrastructure) to provide deep, real-time visibility into complex multi-agent environments. By leveraging a high-performance stack comprising Bun, Express, and Vue, it captures the operational pulse of your AI system without the overhead of heavy persistent databases. It is specifically tailored for developers and power users who need to maintain control over ephemeral agent sessions, tool execution chains, and hook-driven event flows. When agents are running autonomously, this dashboard acts as your window into their decision-making processes, logging every significant interaction from session initialization to completion.

• Real-time WebSocket streaming provides instant updates on agent activities, ensuring you never miss a critical transition or tool call execution.

• Seamless file-system observation monitors the primary raw-outputs event logs, automatically tailing and displaying session timelines and traces as they are written.

• Lightweight and ephemeral architecture avoids persistence layers, performing entirely in-memory to ensure rapid startup and minimal impact on system performance.

• Comprehensive status management through a dedicated control script allows for one-command start, stop, restart, and health-checking of the backend server and client UI.

• Integrated hook support ensures that all internal system events logged by the PAI hook framework are surfaced in an intuitive browser-based interface at http://localhost:5172.

• Use this skill when you need to debug stuck workflows, identify looping tool calls, or simply monitor the efficiency of agent collaborations in real-time.

• Expect the system to read JSONL files from ~/.claude/history/raw-outputs/ and broadcast event data via the server on port 4000.

• Ensure your PAI hooks are configured to capture events into the daily raw-outputs file; the skill relies on this log stream to populate the dashboard.

• Ideal for environments running multiple concurrent agents where understanding the sequence of operations or identifying bottlenecked tool calls is essential for successful system optimization.