Release/v2.1.2

[digitallysavvy]

Summary

• Core SDK update — new generated types and API additions from the June 4 Fern generation
• AgentKit bug fixes — corrects several serialization and precedence issues identified in post-request audit
• New vendors — XaiGrok (xAI MLLM), GenericAvatar, AnamAvatar
• New builder method — Agent.with_audio_scenario()
• New field — interaction_language on Deepgram, Speechmatics, and Sarvam STT options

AgentKit fixes

Breaking change — OpenAI STT (OpenAISTT):
to_config() now validates that model, prompt, and language are all present in the effective input_audio_transcription object. The default model changes from
whisper-1 to gpt-4o-mini-transcribe. Callers who only set api_key will raise ValueError at config-build time; add prompt and language (or supply them via
input_audio_transcription).

TurnDetectionLanguage — removed bare "en" from the TurnDetectionLanguage literal and TURN_DETECTION_LANGUAGE_VALUES tuple; it is not a valid AsrLanguage enum value
and would be rejected by the backend.

VertexAI LLM URL — VertexAILLM.to_config() now constructs the correct Vertex AI endpoint URL from location, project_id, and model instead of falling back to the
Gemini URL.

VertexAI MLLM params — VertexAI.to_config() now places project_id, location, and adc_credentials_string inside params (where the backend expects them) rather than
at the top level of the config map.

LLM field precedence — _resolve_llm_config (agent) and the session's LLM merge now use not in key-existence checks so vendor-configured fields win over the deprecated
agent-level convenience setters.

OpenAI TTS validator — OpenAITTSOptions._validate_byok_params now uses not value instead of value is None so empty-string base_url and model are rejected the same
way Go rejects them.

Deepgram keyterm — DeepgramSTTOptions gains a keyterm field serialized as asr.params.keyterm.

MiniMax preset model stripping (OV-1) — MiniMaxTTS.to_config() no longer includes model, key, group_id, or url in params for the managed-preset path. A
_minimax_preset_model top-level hint enables preset inference and is stripped by strip_inferred_preset_fields before the POST body is finalized.

Test plan

• pytest tests/custom/ passes (82 passed, 1 skipped)
• test_vendor_config_takes_priority_over_agent_level_convenience_fields confirms vendor wins
• test_stt_vendor_params_match_documented_shapes confirms OpenAISTT raises without prompt/language
• test_xai_grok_serializes_v27_shape_without_style confirms new XaiGrok vendor
• test_generic_avatar_omits_session_enriched_fields_when_unset confirms new GenericAvatar vendor