v0.1.6

Highlights

• Voice model picker (#39, closes #33): start_thread accepts a spoken provider and model — the backend resolves wording like "GPT-5.6 Sol" against the provider's model catalog and fails with actionable errors. New list_providers (harnesses + exact model ids) and set_thread_model (change an existing thread's sticky model without starting a turn).
• Threads without a project (#38, closes #26): starting a thread in the personal project ("Don't work in a project") now works — personal workspaces, on any connected machine. Non-git directory projects fall back to working directly in the project's source directory instead of failing on a worktree.
• gpt-live-1 support (#37): OpenAI full-duplex Live API as a voice-session model, with API-key capture and auth-mechanism memory.
• Mobile companion drawer (#27): show threads beside a live call on mobile, with drawer view management, while desktop keeps normal navigation.
• Machine names: start_thread accepts a machine's spoken name, not just a host id; unknown machines list the connected machines instead of HTTP 404.
• Thread outcomes (#35): failures and interruptions are reported instead of guessed at; silent tool lookups (#36) stop spoken preambles before thread reads.

Full changelog: v0.1.5...v0.1.6