Add voice-only Textual TUI for live agent cascade sessions

[alexkroman]

Introduces LiveAgentApp, a hands-free voice-only Textual UI for assembly live (the agent cascade), complementing the existing code-agent TUI. The new TUI drops the text prompt entirely since input is purely voice-based, while reusing the shared chrome (ASSEMBLY wordmark, animated voice bar, transcript message widgets).

Key changes

• New aai_cli/agent_cascade/tui.py: Implements LiveAgentApp (the Textual app) and _TuiRenderer (marshals cascade callbacks onto the UI thread). The app displays a scrolling transcript above an animated voice bar that tracks the session phase (listening/thinking/speaking). A worker thread drives the blocking cascade via run_conversation, with on_stop closing the audio to unblock that worker on quit.

• TUI selection logic in aai_cli/commands/agent_cascade/_exec.py:

  • _should_use_tui(): Determines when to use the TUI (interactive mic session in human mode on a TTY, not file/sample input, not --json/-o text, not piped stdin/stdout).
  • _run_live_tui(): Wires the duplex audio, cascade dependencies, and hands the app a run_conversation closure plus on_stop callback.
  • Refactored _web_search_note() to return the notice string (or None) so it can be passed to the TUI as a notification and still emitted to stderr in line-renderer mode.

• Shared voice-bar infrastructure in aai_cli/code_agent/tui_status.py:

  • Extracted VOICE_FRAMES (animated meter) and _VOICE_PHASES (phase labels + colors) as module-level constants so both the code and live TUIs use identical animations and styling.
  • Added voicebar_markup() helper to generate the voice bar's content (meter, label, accent color, optional hint).

• Message widget enhancements in aai_cli/code_agent/messages.py:

  • Added UserMessage.set_text() to update the user prompt in place (used by the live TUI to grow interim voice transcripts without mounting new widgets).
  • Extracted _user_markup() helper so the styling is consistent between constructor and set_text().

• Comprehensive test suite in tests/test_live_tui.py (342 lines):

  • Direct UI-thread tests for splash, user partial/final, agent replies, interruption, voice-bar animation, and error handling.
  • Worker-thread integration tests driving the real _TuiRenderer through a scripted cascade to cover the off-thread hop, error path, and teardown without a mic/speaker/socket.
  • Integration tests verifying TUI selection logic and wiring into run_agent_cascade.

• Test updates in tests/test_code_messages.py and tests/test_code_tui_status.py: Added tests for UserMessage.set_text() and voicebar_markup().

Implementation details

• The cascade runs on a worker thread; every renderer call hops back to the UI thread via call_from_thread(). Once the app tears down (quit mid-turn), that call raises RuntimeError — the event is moot, so it's dropped rather than surfaced as an unhandled exception.
• The voice bar animates via a 0.3s interval timer, cycling through 6 block-character frames for a smooth pulse effect.
• Quit (Ctrl-C or Ctrl-Q) calls on_stop to close the audio, which ends the mic iterator and unblocks the cascade worker, then exits the app.
• The TUI is disabled for file/sample input, machine output modes (--json, -o text), and non-TTY environments — all fall back to the line renderer.

https://claude.ai/code/session_01WfPLNLa2h4B5khteUmYnFc