Guards

A step debugger for the agent loop: pause at prompts, model requests, responses, tool calls and results; inspect and edit them; play, step or stop.

A plant in the band above the prompt that grows with every tool call and finished turn, wilts when a tool fails, and keeps growing across sessions.

A /trace pane with two tabs: the skills that loaded this session (when, from where, who invoked them, how large) and the web pages Claude fetched or saw in search results, ready to insert as citations