v0.9.160

Patch Changes

• #930 a90258e Three places where Rove decided something was ready before it was.

  rove api add on a repo with a .rove/init.sh no longer reports
  SESSION_FAILED — quoting bun install's progress bar as the reason — for a
  task that goes on to start normally. Repo init records its exit code in a
  per-worktree marker; a marker left empty by an older release meant "re-run
  init" to the launch shell and "init already finished" to the CLI, so the CLI
  spent its engine probe against a shell that was still installing dependencies.
  Both sides now read the marker the same way.

  rove api read-output --source terminal without --tab finds an engine that
  is running on a tab other than tab-1 when the task launches through a wrapper
  command (claudecpa, any custom preset). It was searching the live sessions
  for the task's vendor binary rather than the task's own, and answered "no
  live terminal session for this task" while --tab tab-2 returned a live tail.

  The daemon no longer answers Cannot access 'handlers' before initialization
  to a client that connects during the last moments of its startup — the socket
  now starts accepting only once the request path is complete. This showed up as
  Rove exiting 1 on roughly one launch in five on a busy machine. — @Sma1lboy

• #931 26f8c1a Migrating a legacy kobe install to the Rove layout now succeeds on Windows.

  Every launch reported state migration will retry: state.json: EPERM: operation not permitted, fsync and then migrated nothing, so a machine
  upgrading from kobe kept its old settings, themes and attachments stranded
  under ~/.kobe and ~/.config/kobe forever — the failure left the completion
  marker unwritten, which is what makes the next launch retry, so the warning
  repeated indefinitely.

  The migration flushes each copied file before publishing it, and it was
  reopening the file read-only to do so. Windows backs fsync with
  FlushFileBuffers, which requires a writable handle and rejects a read-only
  one; POSIX flushes an O_RDONLY descriptor without complaint, so the bug was
  invisible everywhere CI runs. Because copyFileSync also carries the source
  file's mode onto the copy, a legacy file with no write bit defeats a writable
  reopen just as thoroughly — on Windows and POSIX alike. The flush now widens
  the temporary file's mode when it has to, then restores the mode it publishes. — @Sma1lboy