Support per-channel WAV files for --system-audio with --save-dir

[alexkroman]

Summary

Extends --save-dir to support --system-audio by writing separate WAV files for each channel (microphone and system audio) instead of rejecting the combination. The microphone stream saves to <stem>-you.wav and system audio to <stem>-system.wav, with a single shared transcript carrying both channels' turns.

Key Changes

• New SaveTargets dataclass (aai_cli/commands/stream/_exec.py): Replaces the tuple return from _resolve_save_targets() with a frozen carrier that distinguishes between single-source audio (audio) and parallel-stream audio (audio_by_label). This allows the session to route each channel to its own file.

• Per-channel filename support (aai_cli/streaming/naming.py): Added channel_audio() function to insert channel suffixes (-you, -system) into auto-named WAV paths while preserving the date bucket and .wav extension.

• Session routing logic (aai_cli/streaming/session.py): New _audio_target() method determines which WAV file a source tees to based on its label and the save configuration. Under --system-audio, each channel looks up its path in save_audio_by_label; otherwise, all sources use the single save_audio path.

• Validation relaxation: Removed the blanket rejection of --save-dir with --system-audio. The explicit --save-audio flag is still rejected for two-stream scenarios (can't merge two streams into one file), but --system-audio-only now works with --save-audio since it's a single stream.

• Test coverage: Added two integration tests demonstrating the feature:

  • test_stream_system_audio_save_dir_writes_one_wav_per_channel: Verifies mic and system streams write to separate WAVs with a shared transcript
  • test_stream_system_audio_only_save_dir_writes_one_labeled_wav: Confirms system-audio-only uses a labeled WAV and never opens the microphone
  • test_save_audio_allows_system_audio_only: Validates that explicit --save-audio works with --system-audio-only

• Documentation updates: Clarified help text and reference docs to explain the per-channel WAV behavior.

Implementation Details

• Uses a frozen SaveTargets dataclass (like StreamOptions) to prevent accidental mid-run retargeting of save destinations.
• Maintains backward compatibility: single-source streams (mic, file, URL, stdin) continue to use the existing save_audio path.
• The shared transcript under --system-audio includes both channels' turns, each prefixed with the source label ("You:", "System:").
• Datetime mocking in tests uses a custom _FixedDatetime class to ensure deterministic auto-generated filenames.

https://claude.ai/code/session_01DKJdMMr8hdLEiBnZWS9A3R