[johnmatthewtennant]

Summary

Berd lets you tune the Spokesperson's realtime session under Advanced: the model, turn detection and eagerness, noise reduction, VAD timing, transcription and the rest. A standalone Berd Call couldn't set any of these, so it always used the defaults.

berd-call start and session now take --spokesperson-options JSON:

• It uses the same camelCase fields and ranges as Berd's settings, for example {"turnDetection":"semantic_vad","eagerness":"low","noiseReduction":"near_field"}.
• It refuses unknown fields, and it refuses the voice and speed, which already have their own flags.
• The options apply when the Spokesperson's session starts. Fields left out keep their current defaults, including any environment overrides for the model.
• Turning the Spokesperson off and on from the menu bar keeps them for the call.
• Saved settings don't keep them. The caller owns them and passes them on each start, so clearing a setting there takes effect instead of reviving an old saved value.

Related issue

None.

Testing

A new test parses a berd-call session command line carrying the options and builds the Spokesperson's session.update from it, the same path a call takes when its Spokesperson session starts:

• The update has semantic VAD with low eagerness, near-field noise reduction, a 512-token limit, the transcription hints and the chosen model. The transcription model and voice keep their configured values.
• Unknown fields, the voice and speed, values outside Berd's ranges, a blank model and malformed JSON are each refused with the field named.
• The arguments kept for saved settings drop the options, while the arguments used to turn the Spokesperson back on keep them.

The released berd-call rejects --spokesperson-options as an unknown argument. No live call was made with the flag.

[johnmatthewtennant]

🤖 Folded into #416 for review and merge.