# M22: many pi windows, frequent switching

Reported with pi run in several herdr panes, switching between background sessions often:

1. a session freezes after switching: its screen stops and keys do nothing, while the
   agent view still shows it running;
2. rows stay stale: "needs input" long after the session went on in another window, a
   session moved to the background stuck on "starting…";
3. the model column shows `anthropic::claude…`, not the part that tells models apart.

## Causes

- **Ctrl+Z reaching the worker.** pi's Ctrl+Z stops its TUI, sends SIGTSTP to its process
  group and restarts the TUI on SIGCONT. A worker leads its own session in the PTY: its
  group is orphaned and the kernel discards SIGTSTP, so the TUI stays stopped for good
  while the agent loop runs on and reports. The client only took a chunk that was exactly
  `\x1a` or `\x1b[122;5u` as its detach key: Ctrl+Z among other bytes (fast typing, a
  multiplexer that batches input) or with lock bits (`\x1b[122;133u`, Num Lock) went on.
- **herdr's pane identity inherited.** M20 carries the whole environment, `HERDR_ENV`,
  `HERDR_SOCKET_PATH` and `HERDR_PANE_ID` included: herdr's pi integration in a background
  worker reported that session's state into the pane the hub ran in.
- **One session, several jobs.** A session moved to the background again (from a pi in
  another window) got a new job beside its old, ended one (`fix-login`, `fix-login (2)`);
  the old row kept its last state, "needs input" included. A session reopened in another
  pi window left its ended job asking for input.
- **No push before the first report.** A pi row was pushed on reports only: until the
  worker's first report it showed "starting…".
- **Lost pushes.** A `list` reply and an event in the same chunk: the event was applied
  before the code after the await, which then replaced the rows with the older list.

## Changes

- `lib/attach.ts` `localDetachAt`: Ctrl+Z anywhere in a chunk, legacy, kitty CSI-u (any
  alternate keys, Caps/Num Lock, text, press or repeat), modifyOtherKeys; not inside a
  bracketed paste. The bytes before it still go to the session.
- worker: Ctrl+Z detaches (pi never suspends in the background).
- host: SIGCONT to the worker's group on every attach (a session stuck this way resumes);
  an attach taken over no longer types or resizes; `terminate-idle` (the reaper's and
  auto-close's stop) decides in the host whether someone is attached; a HELLO while the
  host is finishing gets `exit`; a mirror write that throws does not drop the client's
  output.
- daemon: `supersedeOthers` (a job on a session takes over the rows of the ended jobs on
  it, their name and pin; their records stay), `continuedElsewhere` (a foreground pi
  registering a session turns its ended job's "needs input" or failure into done), a push
  when a live job's worker pid appears.
- `HERDR_` joins the never-carried prefixes.
- agent view: one attach at a time; a 15 s resync; events after a list reply are replayed
  on top of it (`DaemonClient.requestAt`, `position`); listener errors no longer drop the
  rest of a chunk; `ensureLive`'s respawn times out after 20 s.
- Reporter: a failed request backs off (0.5 s doubling to 5 s) even for state changes.
- `shortModel` drops `vendor::` and Bedrock prefixes and `-vN:M` suffixes; `fitTail`.

## Tests

- `test/stress-attach.test.ts`: 60 rounds (600 with `STRESS=1`) of attach, takeover,
  abrupt drops, resize storms, two attaches at once and clients that stop reading against
  a worker that streams, echoes and handles SIGUSR2; a flood into a client that never
  reads; detach/attach bursts; one PTY switching between two sessions with Ctrl+Z, keys
  typed before and with it (40, 300 with `STRESS=1`); a worker that suspends itself as pi
  does is resumed by the next attach (fails without the host's SIGCONT).
  - M42: that last test failed now and then under a full `npm test`. The fake worker
    dropped keys while "stopped", and the host's SIGCONT and the keys typed right after
    the attach reach a node process through the same event loop in no fixed order; pi
    stops reading its terminal instead, so keys typed meanwhile wait in the terminal.
    The fake worker now does the same (pauses stdin until SIGCONT), the test waits for
    the ticks to stop rather than a fixed 300 ms, and a timeout on a PTY's output says
    what the terminal showed last.
- daemon: rows taken over, continued elsewhere, pid pushed before a report.
- reporter back-off, client stream order, detach key encodings, model names, herdr env.