Read what a session is doing

A server-sent event stream. Each transcript entry is an `entry` event with `id: <seq>`; approvals, questions, waiting messages, status and usage arrive as `approval`, `ask`, `pending`, `status` and `usage` events, each `data` a `SessionEvent`. While the model writes, `reply` events carry its text as it streams (an id, an offset and the text from there; offset 0 is the whole text so far, which a stream opened mid-reply gets first): live only, never replayed, ended by the assistant entry itself. Reconnect with `Last-Event-ID` (or `?after=`) to get every later entry exactly once. A stream that has to end says why in an `error` event.