v4.100.259

Agents

• Every guard says when it acts, in the chat and on the event stream.
  Loop detection, which catches a model sending the same tool call over and
  over, gave its warnings inside the tool result and stopped the run through
  the mistake count, so neither the chat nor an SDK client could tell a loop
  from any other run that ended early. It now posts a status line for each
  step: the first warning, the last warning, and the stop. A request to check
  edited files before finishing, a reminder about open checklist items and a
  reminder about delegated work that is still running each post one too.

  For SDK clients, each of these is a status notice (type: "notice",
  noticeType: "status") with metadata.kind naming the guard (loop_detection, edit_verification,
  checklist_close_out, team_obligation) and metadata.phase saying what it
  did (soft, hard and stopped for loop detection, nudge for the
  others). A cut loop keeps the kind reasoning_loop and now says in
  metadata.cause whether the reasoning, the reply or a tool call's arguments
  looped, and a removed block of repeated reasoning says in
  metadata.nudged whether the model was also asked to change approach.