feat!: sdk regeneration 2026-04-14

[GregHolmes]

Summary

• Fern SDK regeneration with latest API specs from deepgram-docs
• Re-applied manual patches (broad exception catch, optional send params, sanitize numeric types, construct_type keyword args, float→int
  type fixes)
• Fixed examples for current API surface
• New voice agent configurations and variables REST API
• New WebSocket methods: send_update_think (agent), send_configure (listen v2)

BREAKING CHANGE

Removed public type exports due to upstream API spec restructuring:

Agent provider types restructured (~280 types):

• AgentV1Settings*SpeakEndpoint* and AgentV1Settings*SpeakOneItem* types replaced by shared provider schemas
  (SpeakSettingsV1Provider_*)
• AgentV1UpdateSpeak* types restructured similarly
• AgentV1Settings*ThinkOneItem* types replaced by shared ThinkSettingsV1 schemas

Type enums inlined to Literal strings (42 types):

• ListenV1MetadataType, ListenV1ResultsType, ListenV1SpeechStartedType, ListenV1UtteranceEndType
• ListenV2ConnectedType, ListenV2FatalErrorType, ListenV2TurnInfoType
• SpeakV1MetadataType, SpeakV1TextType, SpeakV1WarningType
• AgentV1*Type enums (29 types)

Other removed types (9 types):

• AgentThinkModelsV1ResponseModelsItem*Provider (6 types)
• MediaTranscribeRequest*Zero (3 types)

Not breaking:

• No client methods removed
• No method signatures changed
• Core APIs (listen, speak, manage, read) unchanged
• DeepgramClient, transcribe_url, transcribe_file, audio.generate all work identically

Test plan

• Unit tests: 147 passed, 1 skipped, 0 failed
• Integration tests: 32/32 passed
• Examples: all passing
• 6 SDK-based starters validated against local branch