Monitors
Interactive tabs over env_habitat, outside the graph engine
Membership test: does this page document a top-level frontend tab backed by a lifespan-owned runner service that drives env_habitat directly — no graph execution in the loop? Both members spawn their own auto_host through BaseServer, talk to it over the same HTTP surface the agent drivers use, and score with habitat's own env_habitat__evaluate — the same ruler as the verified graphs. The experiment these tabs serve is described in the Coding-Agent Harness capability page.
Coding-Agent Monitor
One-click host for a Claude-SDK run + log browser over all three harness output roots — artifact-derived state, EventSink vocabulary, inline frames
Human Performance Tab
A human in the agent's seat — one keypress = one discrete action, same episodes, same metrics ruler, records under outputs/human/