v1.7.1

Two process-global stores in the CLI leaked between the sessions one process
serves. A patch release, so nothing here changes the protocol surface 1.7.0
introduced.

Fixed

• A todo list cannot leak into another thread. tools.rs kept the list in
  a single process-global slot, shared by every session the process serves, so
  a plan written in one conversation showed up in the next. It is keyed by
  session now, like the permission grants, and cleared when the session ends.
  The TUI carries a session key of its own (permissions::TUI_SESSION) rather
  than the session-less "" key, so its background commands, plan and
  permission grants are torn down on exit; the unused TODOS_TEST_LOCK and
  reset_todos are gone.
• The model picker's roots are resolved per caller instead of through the
  environment. setup.rs built its model store by mutating HF_HUB_CACHE,
  HF_HOME and HOME under ENV_TEST_LOCK, which never isolated the values
  because the picker reads those variables on other threads — setup::tests
  failed intermittently. Production callers resolve the roots once with
  ModelRoots::from_env() and pass them into the internals, so a test points
  the readers at a scratch directory instead of aiming process-global
  variables at it.
• sigit publishes again. The crate listed eight keywords and crates.io
  accepts five, so the v1.7.0 publish failed with a 400. The list is trimmed to
  cli, ai, coding-agent, llm, acp.

---

Full changelog: https://github.com/getsigit/sigit/blob/v1.7.1/CHANGELOG.md

---

About siGit Code: sigit.si/code