v0.9.147

Patch Changes

• #897 db1eb2b The state directory and the plugin tree are repaired to owner-only on every
  start, not just when they are created

  <home>/.rove was created by a bare mkdir and landed at 0755 under the
  default umask, and the daemon socket inside it bound at srwxr-xr-x because
  listen() applies the umask too. Nothing else gates that socket — connections
  are accepted with no peer-credential check — so on a shared machine the
  directory mode was the entire ACL, and reaching the socket means add (launch
  an engine) and send (text into a live session). Both the daemon and the PTY
  host now create that directory 0700 and chmod it again on every boot, and both
  sockets are chmod'd 0600 after the bind.

  The "again on every boot" half is the one that matters. mkdirSync and
  writeFileSync apply their mode only when they create the path, so a home
  that predates the mode arguments kept 0755 forever — the population a
  creation-time fix cannot reach is exactly the exposed one.

  The same defect had the plugin tree world-readable.
  docs/PLUGIN-AUTHORING.md tells authors to keep API keys in the config .env
  and states that the .env, the state directory and log.jsonl are owner-only;
  that was true for a fresh install and false for every older one, because
  writePluginSettings rewrites the .env in place and an in-place write never
  changes a mode. Every registered plugin's config, state, .env, log.jsonl
  and the registry are now repaired at daemon start and after each settings save,
  so the documented sentence holds on installs that already exist.

  Two comments in the web-token chain described mechanisms that are gone.
  pty-auth.mjs credited the daemon with minting the token file it fails closed
  without — no daemon path does; kobe-harness/dev.ts is the only minter, which
  matters to anyone adding a launcher. The harness token header was written
  around a <meta name="rove-web-token"> injector that #855 deleted along with
  the daemon-hosted web transport, and presented the only channel that still
  works, VITE_ROVE_WEB_TOKEN, as a narrow fallback. Both now say what the code
  does. — @Sma1lboy

• #898 ca20e8f Fix the two commit-tree calls in the daemon test fixtures that never passed a git identity, which failed every release run on CI where no global user.name exists. — @Sma1lboy