Endpoint capability companions

C9-A adds pure getRealtimeModelCapabilities(query) and getTranscriptionCapabilities(query) on root and oneringai/shared. They do no network IO and import no provider SDK or Node runtime in browser bundles. They do not replace model/STT registries, infer vendor from model text or change legacy Realtime validation, settings, registry objects or STT streaming behavior.

typescript

import { Vendor, getRealtimeModelCapabilities, getTranscriptionCapabilities }
  from 'oneringai/shared';
const voice = getRealtimeModelCapabilities({
  vendor: Vendor.OpenAI, model: 'gpt-realtime-2', transport: 'webrtc',
});
const input = getTranscriptionCapabilities({
  vendor: Vendor.OpenAI, model: 'gpt-transcribe', mode: 'conversation-input',
  transport: 'websocket', conversationModel: 'gpt-realtime-2.1',
});

Reasoning support and exact allowed efforts are separate. support distinguishes supported/unsupported/unknown. allowedEfforts: null explicitly means the exact set is unverified; an empty array means reasoning is explicitly unsupported. Existing saved efforts remain host-owned and visible even when absent from a verified set. Metadata does not silently migrate settings or enforce a stronger validator.

For the official endpoint, GPT-Realtime-2 has the documented minimal, low, medium, high, xhigh set. The 2.1/2.1-mini model pages establish reasoning but not their exact effort sets, so those sets remain unknown. Provenance distinguishes checked documentation from projected existing registry declarations. OpenAI reasoning guide, 2.1, 2.1-mini. Facts were checked on 2026-10-28; no live model probes were run.

Transcription modes are file, transcription-session, and conversation-input. File support projects existing declared audio-transcription endpoints. Dedicated gpt-transcribe committed-turn sessions require WebSocket; WebRTC/SIP are explicitly unsupported for that documented workflow. Dedicated gpt-live-transcribe supports WebSocket/WebRTC with VAD omitted or off; server_vad and semantic_vad are unsupported. Exact conversation/transcription model pairings stay unknown when unverified, including the example above. A required transport does not prove an unlisted pairing works. OpenAI transcription guide.

Unknown models, vendors and custom endpoints stay unknown. Set the query's endpoint: 'custom' for a custom gateway even if its model name matches an OpenAI model. Defaults describe the named vendor's official endpoint. There is no automatic network research, Cartesian product table or generic capability-rule language.

Run npm run example:endpoint-capabilities for all three support states and the known-versus-unknown effort distinction. AMOS has no Realtime selector and receives no voice UI here. ICOS must explicitly adopt these lookups while preserving saved values and its host authorization/custom-service rules; an SDK pin alone does not change selector choices. These lookups are included in oneringai@2.0.0-beta.1; publication does not imply a consumer upgrade.