@@ -5,6 +5,13 @@ description: Generate AI voiceovers, sound effects, and music using ElevenLabs A
 
 # ElevenLabs Audio Generation
 
+## Current v4 transports
+
+For v4, v4 Turbo, Instant Soundtracks and realtime models, read
+[the October integration guide](../provider-refresh-oct23/SKILL.md).
+v4 uses Text to Dialogue; Turbo uses the dialogue WebSocket.
+Do not send those models to the legacy text-to-speech endpoint.
+
 ## OpenMontage provider routing
 
 Inspect the OpenMontage registry before choosing an authentication path.
 
@@ -103,3 +103,10 @@ Contracts were inspected 2026-10-03; each schema records its official source.
 See docs/provider-update-plan-2026-10-03.md for release research. Public Kling 4
 generation endpoints remain unverified, so do not invent a model ID. Live voice
 agents require a separate streaming/session integration beyond batch narration.
+
+## October 23 follow-up
+
+For current Google retirement handling, Eleven v4 transports, and newly supported
+providers, read [provider-refresh-oct23](../provider-refresh-oct23/SKILL.md).
+The v4 narration HTTP payload is dialogue inputs/settings, not legacy TTS
+text/voice_settings; v4 Turbo is WebSocket-only in the direct file adapter.
@@ -0,0 +1,136 @@
+---
+name: provider-refresh-oct23
+description: API contracts and production routing for the October 23, 2026 provider refresh.
+---
+
+# October 23 provider integrations
+
+Read the tool's current input schema before use. Announce tool, host, exact model,
+operation and quoted cost before production. No automatic host/model fallback.
+Use explicit output paths under the production project. Account access is not
+established merely by having a credential configured. Most new prices require a
+fresh account quote; do not use expired promotions as permanent prices.
+
+## Google and ElevenLabs migrations
+
+- `gemini_omni_video`: direct Gemini Interactions API model is
+  `gemini-omni-1.1-flash`. Old preview retired October 22. Continue editing with
+  `previous_interaction_id`; do not substitute it inside a Veo request.
+- `veo_video`: Google backend requires Vertex; fal retains its own contracts.
+  Gemini API Veo has retired. Select Omni explicitly for the replacement workflow.
+- `google_imagen`: new default `gemini-nano-banana-2.1`; Imagen models require
+  Vertex. Native Gemini image models use generate_content, not Imagen predict.
+- `elevenlabs_tts`: `eleven_v4` uses HTTP Text to Dialogue with `inputs` and
+  `settings`. Keep each narration request <=2,000 characters. `next_text` maps to
+  `future_text`; context strings <=100 characters. Style/speed/speaker boost are
+  not dialogue settings. `eleven_v4_turbo` uses the dialogue WebSocket; this file
+  adapter does not support timestamps or continuity controls on Turbo. It waits
+  for global `is_final`, not the per-turn marker. The current official realtime
+  tutorial documents Turbo although the older WebSocket reference says v3-only.
+
+## Narration and conversion
+
+- `heygen_tts`: model_id `heygen-voice-1`, `heygen-voice-1-turbo`, or
+  `heygen-voice-1-turbo-free`. Pass an ACTIVE native public/workspace voice ID;
+  `operation=list_voices` discovers native voices. Professional clones only work
+  with the non-Turbo model. `voice_type` can prevalidate the category; the service
+  validates actual ownership/status. Completed speech and SSE both write WAV.
+  SSE parts are individual WAV containers, decoded before assembly. Timestamps
+  become an alignment sidecar and word_timestamps. Free access ends November 30,
+  2026; failures never switch to paid. Rate rejection retries are bounded and honor
+  short Retry-After delays; larger delays surface to the caller.
+- `heygen_catalog_tts`: separate catalog IDs, `/v3/voices/speech`; explicitly choose
+  `engine` (starfish, heygen_voice, elevenlabs). Native professional clone IDs do
+  not belong here. Catalog output can be MP3 or WAV. Supply either suffix; the returned artifact
+  path uses the actual response codec suffix and preserves the audio.
+- `heygen_voice_convert`: ACTIVE instant workspace voice only, at most300 seconds.
+  Exactly one audio_path (<=16 MB), audio_url or asset_id. Output WAV. Supply a
+  stable idempotency_key if resuming after an ambiguous timeout; no automatic retry.
+- `grok_tts`: xAI narration `/v1/tts`, not Grok Voice realtime. `language=auto` by
+  default, voice_id defaults to eve; list_voices discovers voices. output_format
+  is an object, e.g. {"codec":"wav","sample_rate":24000}; this file adapter exposes
+  MP3/WAV, not headerless telephony codecs. Optional character alignment is saved.
+- `mistral_tts`: model_id `voxtral-mini-tts-3-0`; exactly one voice_id or local
+  reference_audio_path. mp3/wav/flac/opus. Inline emotion/pace tags and expressiveness;
+  free-form style instructions are not available at launch. No realtime TTS claim.
+- `mai_tts`: Azure Speech key/region, model_id MAI-Voice-2.1 or MAI-Voice-2.1-Flash.
+  voice_id is a managed voice base such as en-US-Harper (the model suffix is added).
+  The TTS selector translates speed and pitch through the Azure SSML adapter.
+  This adapter does not provision personal voice clones or grant gated access.
+- `mistral_stt`: Voxtral Mini/Small Transcribe3, local audio files; timestamps are
+  segments, not words. Diarization implies segment timestamps. Raw transcript JSON
+  is persisted, including speaker and usage metadata.
+- `elevenlabs_soundtrack`: video_paths (<=10, combined <=200 MB / 600seconds),
+  optional description/tags, add_sfx/use_audio_as_reference/separate_tracks. Separate
+  tracks return playable files in a sibling *_tracks directory; never treat a ZIP
+  as an audio file. The adapter rejects unsafe archive paths and oversized archives.
+
+## Images and video
+
+| Model | Explicit tools / hosts |
+|---|---|
+| Nano Banana2.1 | google_imagen, gemini_fal_image, gemini_replicate_image |
+| FLUX3 Image | bfl_image, flux3_fal_image, flux3_replicate_image |
+| FLUX3 Max video | flux3_video (fal) |
+| Vidu Q4 | vidu_video (fal), image_to_video or reference_to_video; no text-only route |
+| Fibo2 | bria_image direct or bria_fal_image |
+| MAI Image2.6 / Flash | mai_image, deployed Foundry models |
+| Grok Imagine1.5 / Lite | grok_video direct xAI |
+
+fal and Replicate have checked-in endpoint schemas under tools/provider_contracts.
+Pass provider_params only when documented by that host. Do not copy fal fields to
+Replicate or direct APIs. Existing Pro image model defaults remain unchanged on
+aggregators; select the new model explicitly.
+
+BFL bounding-box layouts are encoded in the prompt per BFL docs, not a fabricated
+`boxes` endpoint field. BFL supports up to10 reference images; width/height must be
+provided together. Its API reference is authoritative over older narrative examples.
+Bria persists structured metadata for later editing. Its shared endpoint can route
+some legacy combinations to older Fibo models; the direct tool rejects those
+combinations when advertising Fibo2. MAI requires model identity to match the
+configured deployment; PNG output and <=2,359,296 pixels with each side >=768.
+
+Grok1.5 references are limited to720p. Lite has no reference audio, reference images,
+last_frame or keyframes. Full1.5 supports preset voice references; custom audio
+references remain partner gated. Mid-video pins must be inside duration and occupy
+different1/3-second slots. No model substitution on failure. fal/Replicate/BFL/Grok
+jobs expose resume metadata; save job_path before polling and resume pending jobs
+instead of submitting duplicate paid generations.
+
+## Realtime is separate from narration
+
+`realtime_audio` is a bounded one-response file runner for Grok Voice3,
+Eleven Realtime Mini/Max and MAI-Transcribe-2-Streaming. Input is text or mono PCM16
+WAV (16/24kHz; Eleven requires24kHz). Voice output is WAV; MAI output is event JSON.
+Input sending and event receiving run concurrently. MAI commits explicitly and
+waits for final transcription, not a commit acknowledgment. Eleven uses ephemeral
+session credentials and bursty input pacing for recorded files.
+
+For live audio, VAD, interruptions, multiple turns, or function calls, use the
+context-managed `tools.audio._realtime_session.RealtimeAudioSession` transport:
+`append_audio`, `commit`, `receive`, `request_response`, `cancel`, and
+`submit_tool_result`. The caller owns the session lifecycle and tool approvals.
+Received tool calls are never executed automatically. The finite file runner
+rejects tools because it cannot continue them after closing the socket. No microphone
+UI, browser app, or voice-agent orchestration is implied by this transport layer.
+Secrets remain in memory. No realtime translation or Kling4 support was added.
+
+## Primary references
+
+- https://ai.google.dev/gemini-api/docs/deprecations
+- https://elevenlabs.io/docs/api-reference/text-to-dialogue/convert
+- https://elevenlabs.io/docs/eleven-api/guides/how-to/websockets/realtime-tdd
+- https://elevenlabs.io/docs/api-reference/music/video-to-music
+- https://developers.heygen.com/docs/voices/heygen-voice-speech
+- https://developers.heygen.com/docs/voices/speech-to-speech
+- https://developers.heygen.com/reference/generate-speech
+- https://docs.x.ai/developers/model-capabilities/audio/text-to-speech
+- https://docs.x.ai/developers/model-capabilities/video/overview
+- https://docs.mistral.ai/studio/audio/overview
+- https://docs.bfl.ml/api-reference/utility/generate-an-image-with-flux-3
+- https://docs.bria.ai/whats-new/fibo-2
+- https://learn.microsoft.com/azure/ai-services/speech-service/mai-voices
+- https://learn.microsoft.com/azure/foundry/foundry-models/how-to/use-foundry-models-mai-image
+- https://learn.microsoft.com/azure/ai-services/speech-service/mai-transcribe-2-streaming-realtime
+- https://elevenlabs.io/docs/eleven-agents/guides/openai-realtime-compatibility
+- https://docs.x.ai/developers/rest-api-reference/inference/voice