Get agent roster

What each named agent (task_runs.agent_name) has actually done, in one report: runs, minutes and tool calls spent, how many of its tasks a human approved vs rejected, the rejection rate, the impact/value (Task.value) delivered by the approved ones, and tool calls spent per approved task. Runs made with the account's full-access token, and everything predating agent-name attribution, collapse into one agent_name: null bucket rather than being hidden. A decided task with no run of its own in the window isn't guessed at — it's simply not counted. Each entry also carries what that agent holds right now (not windowed), by lane the way next_task resolves it: open (queued, unclaimed), in_progress (under an active claim) and waiting_on_human (tasks at pending_approval/plan_proposed plus open decisions); an agent holding work with no run in the window still appears. Reports only: no score, no gate, nothing written.