v0.12.0 — watch follows the server

Added

• agent-rack watch follows the whole server, not one session. With no sessionId it attaches
  to a running SSE server and streams every sub-agent that runs there — several concurrently, and
  ones started after it — so it can be left running in a spare terminal before anything exists.
  Each line is prefixed <agentId>:<first 8 of sessionId> (--json adds a sessionId field to
  every event instead), because concurrent agents interleave on one stream.
• An empty server is a wait, not an end state. watch keeps listening while idle, both before
  the first session and after the last one finishes; previously it printed No sessions tracked by the agent-rack server at <url> and exited, which is exactly the moment someone who just opened a
  terminal needs it to stay. --exit-when-idle opts out, for a script that wants it to return once
  the runs it followed are done.
• agent-rack watch / agent-rack session watch itself (introduced in 0.11.2, undocumented
  here): tail -f for a background sub-agent. It holds the connection and polls for you, unlike
  the one-shot session status/session tail, advancing by cursor rather than by count — so an
  agent producing output faster than the poll interval cannot outrun it and have events skipped,
  and a real gap is reported ([N event(s) dropped from the retained log]) instead of passed off as
  complete. Ctrl-C closes the connection; the session keeps running.

Changed

• watch with no sessionId no longer means "the newest session". It now means "everything on
  this server", so pickWatchTarget is gone in favour of classifyWatchSessions. Sessions that had
  already finished when watch started are not replayed — a server's whole history would bury the
  run being waited on — so it reports how many it ignored and you can pass one's sessionId to see
  it. A session that starts and finishes inside one poll interval is still shown in full.
• A terminal status line now prints below that session's final events, not above them. Status is
  still read before draining, so a session that settles between the two calls keeps its last
  events; only the closing line moved.
• The plugin's PostToolUse monitor-nudge hook is gone (0.11.2). It asked the model to babysit each
  background session from inside the conversation, which either flooded the context window or went
  silent until the run was over; agent-rack watch puts that output in the user's own terminal for
  free. tool-selection now offers it as the first answer when someone wants to see progress.