v0.9.151

Patch Changes

• #905 310f0a4 Acquire exclusive daemon home ownership before migrations and core initialization, and keep it through shutdown and pending writes. Competing sockets can no longer start two writers for the same home; crashed owners release their lease automatically.

  Clean up orphaned terminal sessions using their observed generation so delayed sweeps cannot kill a newly started session. Keep unknown host or session liveness distinct from confirmed absence, and never borrow a sibling session's completion marker.

  Serialize settings read-modify-write transactions across processes, including corrupt-file recovery, so independent settings changes survive concurrent writers. — @Sma1lboy

• #907 9c57172 Fix five ways a destructive path could lose work without saying so.

  • A salvage snapshot built its throwaway index empty, so git treated tracked
    files as untracked and .gitignore applied to them. Every uncommitted edit to
    a file the repo tracks and its own .gitignore also matches (a committed
    dist/README.md, a committed server.log) was recorded as a deletion while
    the snapshot reported success. The index is now seeded from HEAD.
  • One gitignored file whose name starts with - made du read it as an option,
    which emptied the ignored-work probe for the whole worktree — so the non-force
    delete gate stopped refusing and HANDOFF.md / .scratch/ were destroyed with
    no snapshot. du now gets a -- terminator, batches under ARG_MAX, and
    measures newline-containing names one at a time.
  • Two salvages in the same second wrote the same ref and the second silently
    overwrote the first, though both callers were told their work was saved. The
    write is now create-only, and a non-ASCII branch keeps its name in the ref
    instead of collapsing to detached.
  • With the "next to project" worktree location, an orphaned worktree (upstream
    .git gone) could never be removed: the managed-root guard could not expand
    $project_dir without the repo, so the task parked in deletion.phase: error
    and every retry re-ran the same unsatisfiable branch.
  • rove theme remove did not validate the name, so remove '../../notes'
    deleted any reachable .json. Both add and remove now share one check. — @Sma1lboy

• #909 77e1674 CI finishing on one of your tasks now says so, instead of only changing a chip

  The daemon has polled gh pr list per task for a while, and it already wrote
  checkState onto the task and pushed it to every client. Nothing in the TUI
  ever read that field. So the only signal that a PR's checks had landed was the
  sidebar chip changing colour — which you see if you happen to be looking at
  that row, and which is exactly the wrong shape for the case the poller was
  built for: four tasks in flight, you in the fifth.

  Checks resolving now raise a toast naming the task, whether they passed or
  failed, and which PR. Only the resolution is announced: a run that merely
  started (no checks → pending) stays quiet, as does every flap in between, and a
  task whose checks were already settled when the daemon came back does not
  re-announce itself on restart. The rule was written and unit-tested when the
  poller shipped (checkResolutionNotify); it had no subscriber until now.

  Two call sites that were each deriving "is this task's repo a remote ssh://
  project?" by hand — the engine launch builder and the workspace centre column —
  now ask remoteKeyForRepo, the helper that was already written to be the one
  place that decision is made. — @Sma1lboy

• #908 276c7ce Keep Claude, Codex, and Kimi trust writes and Claude/Codex hooks in the selected vendor profile. Preserve invalid or oversized configuration and private file permissions, recognize Rove hook commands without removing user lookalikes, and keep legacy startup cleanup in the active profile. Stop Codex trust writes when its config lock times out. — @Sma1lboy