v0.9.153

Patch Changes

• #911 2378381 Merge three pieces of logic that were written twice and had drifted apart.

  Worktree adoption ignored your worktree location setting. The daemon
  computed managed worktree roots from its own copy of the layout, and that copy
  never learned about worktree.basePath. If you moved your worktree location in
  Settings → General, an engine starting inside one of those worktrees was not
  recognized, so the worktree never became a task — with no error to explain it.
  Both sides now derive the roots, and the meaning of the stored setting, from
  one module.

  Sidebar +N −M chips counted lines the parser rejects. The daemon's
  counter re-scanned git status output with a looser filter than the shared
  porcelain parser: a status pair with no path, and any line without a separator
  in column 3 (a stray warning, a truncated read), were each billed as one added
  file. The counter is now built on the parser, which is also the one the
  porcelain edge-case tests cover — previously they covered a copy that was no
  longer in this path.

  A whitespace-only .rove/pr-instructions.md blanked the prompt. The
  .rove/ → .kobe/ fallback was implemented three times with two different
  ideas of "non-empty". A file holding just a newline was returned as the PR/CI
  template (producing an empty prompt) while the same file was correctly skipped
  for init-prompt.md. Trimming now applies everywhere: a file of pure
  whitespace is a placeholder, and the next candidate — or the built-in
  template — is used.

  rove repo show said the opposite of what runs. It reported "present
  (wins)" from a bare existence check, so an empty .rove/init-prompt.md was
  shown as winning while the runtime actually used .kobe/ or your saved
  override — in the one situation the command exists to resolve. It now reports
  from the same rules the runtime applies, and distinguishes a file that is
  shadowed by a higher-precedence one from a file that is ignored for being
  empty. — @Sma1lboy