Add --show-code flag to agent-cascade for script generation

[alexkroman]

Adds a --show-code flag to the assembly agent-cascade command that generates and prints the equivalent Python script instead of running an interactive conversation. This mirrors the existing --show-code functionality in the agent command.

Key Changes

• New code generation module (aai_cli/code_gen/agent_cascade.py): Renders a complete, runnable Python script that reproduces the cascade session by wiring Streaming STT → LLM Gateway → streaming TTS client-side. The script:

  • Uses a single full-duplex sounddevice.RawStream for mic capture and speaker playback (required for macOS CoreAudio compatibility)
  • Streams STT at 24 kHz with configurable speech model and turn formatting
  • Calls the LLM Gateway with the running conversation history
  • Synthesizes replies sentence-by-sentence via streaming TTS
  • Supports barge-in (interrupting a reply with a new user turn)

• Static body module (aai_cli/code_gen/agent_cascade_body.py): Separates the orchestration logic (with many literal braces for dict/set literals and protocol loops) from the formatted header, avoiding the need to double-escape braces.

• CLI integration (aai_cli/commands/agent_cascade/_exec.py):

  • Adds show_code: bool option to AgentCascadeOptions
  • Implements _print_show_code() to generate and print the script without authenticating or opening audio
  • Warns on stderr (not stdout) if a file source is passed, since the generated script is microphone-driven

• Public API (aai_cli/code_gen/__init__.py): Exports agent_cascade(config, speech_model) function for script generation.

• Tests:

  • tests/test_code_gen_agent_cascade.py: Example-based tests verifying all three legs (STT/LLM/TTS) are wired, session knobs are injected, and the generated code is valid Python
  • tests/test_agent_cascade_show_code.py: CLI integration tests covering the print-only path, sandbox-only enforcement, and warning behavior

• Documentation: Updated help text and snapshot to reflect the new flag.

Implementation Details

• The generated script targets the active environment's hosts (sandbox-only, since streaming TTS has no production host), so URLs are always valid for the key's environment
• Named per-leg knobs (voice, model, max_tokens, etc.) are reflected; escape hatches (--stt/--llm/--tts-config) are not
• The reply-cue predicate adapts based on format_turns: with formatting on, it waits for the punctuated turn; off, a bare end-of-turn fires
• The script is printed to stdout (clean for > script.py), with warnings on stderr if needed

https://claude.ai/code/session_01Bx8YSbUmmzaQFqyJPRHrq3