@kaana.ai/contracts 0.0.0-stage → 4.3.0

package/dist/public/realtime.js

@@ -0,0 +1,797 @@
+/**
+ * Realtime sessions as a public client speaks them (`GET /v1/realtime`).
+ *
+ * Moved from `@oxy.so/contracts` 4.12.0 (`src/inference/realtime.ts`) by Kaana ADR 0002.
+ */
+import { z } from 'zod';
+import { realtimeAudioFormatSchema } from './audioFormats.js';
+import { realtimeSessionKindSchema, realtimeSessionTransportSchema, } from './catalogue.js';
+import { inferenceTimestampSchema, modelReferenceSchema, PADDED_BASE64_PATTERN, requestIdSchema, } from './identifiers.js';
+import { publicErrorSchema } from './errors.js';
+import { requestLabelsSchema, toolChoiceSchema, toolDefinitionSchema } from './messages.js';
+import { usageQuantitySchema, usageSourceSchema } from './money.js';
+import { inferenceFinishReasonSchema } from './streamEvents.js';
+/**
+ * The most base64 characters one audio frame may carry: 49 152 bytes of audio,
+ * about one second of 24 kHz mono PCM16. Bounded so no frame is large enough to
+ * stall a connection, and so a byte cap can be enforced as audio arrives rather
+ * than after it has been buffered.
+ */
+export const MAX_REALTIME_AUDIO_FRAME_BASE64_LENGTH = 65_536;
+/** The longest a session may be signed for: one hour. */
+export const MAX_REALTIME_SESSION_DURATION_MS = 3_600_000;
+/**
+ * The most audio a session may be signed for in EACH direction: one hour of
+ * 24 kHz mono PCM16 (48 000 bytes per second), the densest format a session
+ * carries.
+ */
+export const MAX_REALTIME_SESSION_AUDIO_BYTES = 172_800_000;
+/** The longest a dropped connection may be resumed after: one minute. */
+export const MAX_REALTIME_RESUME_WINDOW_MS = 60_000;
+/** A client-chosen command identity, unique within one session. */
+export const realtimeCommandIdSchema = z.string().min(1).max(128);
+/** A conversation item's identity within one session. */
+export const realtimeItemIdSchema = z.string().min(1).max(128);
+/** A response's identity within one session. */
+export const realtimeResponseIdSchema = z.string().min(1).max(128);
+/** What a session's responses produce. */
+export const realtimeOutputModalitySchema = z.enum(['text', 'audio']);
+const audioFrameSchema = z
+    .string()
+    .min(4)
+    .max(MAX_REALTIME_AUDIO_FRAME_BASE64_LENGTH)
+    .regex(PADDED_BASE64_PATTERN);
+const sequenceSchema = z.number().int().nonnegative().safe();
+const milliseconds = z.number().int().nonnegative().safe();
+/**
+ * How the end of a user's turn is detected.
+ *
+ * `none` means the client commits input audio itself. The two VAD forms name
+ * `createResponse` and `interruptResponse` explicitly rather than defaulting
+ * them: whether the model answers on its own, and whether new speech cuts it
+ * off, are the two behaviours a caller most needs to have chosen, and a default
+ * would choose them silently — and differently per provider.
+ */
+export const realtimeTurnDetectionSchema = z.discriminatedUnion('type', [
+    z.object({ type: z.literal('none') }).strict(),
+    z
+        .object({
+        type: z.literal('server_vad'),
+        threshold: z.number().min(0).max(1).optional(),
+        prefixPaddingMs: z.number().int().min(0).max(5_000).optional(),
+        silenceDurationMs: z.number().int().min(0).max(10_000).optional(),
+        createResponse: z.boolean(),
+        interruptResponse: z.boolean(),
+    })
+        .strict(),
+    z
+        .object({
+        type: z.literal('semantic_vad'),
+        eagerness: z.enum(['low', 'medium', 'high', 'auto']),
+        createResponse: z.boolean(),
+        interruptResponse: z.boolean(),
+    })
+        .strict(),
+]);
+/**
+ * Transcription of the caller's input audio. Presence enables it; the
+ * transcription model is the deployment's own, declared where the deployment
+ * is, never chosen by a string here.
+ *
+ * In a `conversation` it is served only by a model whose catalogue entry says
+ * so (`capabilities.realtime.inputAudioTranscription`, contract set 4.1.0),
+ * and each transcribed turn arrives as `transcript.delta` / `transcript.done`
+ * with `source: 'input_audio'`. What it costs is settled with the session:
+ * where the provider bills the transcription apart from the session model, as
+ * `input_transcription_milliseconds` (see `USAGE_UNITS`); where the session's
+ * own price includes it, as nothing more.
+ *
+ * `language`, when given, is one of the model's published `audio.languages`.
+ */
+export const realtimeInputTranscriptionSchema = z
+    .object({
+    /** A BCP 47 tag, when the caller knows the language. */
+    language: z.string().min(2).max(35).optional(),
+    prompt: z.string().max(4_096).optional(),
+})
+    .strict();
+/** The target of a `translation` session. */
+export const realtimeTranslationSchema = z
+    .object({
+    /** A BCP 47 tag. */
+    targetLanguage: z.string().min(2).max(35),
+})
+    .strict();
+/**
+ * A session's full configuration, as signed at open and as echoed back by
+ * `session.created` and `session.updated`.
+ *
+ * `inputAudioFormat` and `turnDetection` are required because every session
+ * kind consumes audio and has to know when a turn ends; nothing else is,
+ * because each kind needs a different subset — see the refinement on
+ * {@link realtimeSessionRequestSchema}.
+ */
+export const realtimeSessionConfigSchema = z
+    .object({
+    instructions: z.string().max(32_768).optional(),
+    outputModalities: z.array(realtimeOutputModalitySchema).min(1).max(2).optional(),
+    voice: z.string().min(1).max(64).optional(),
+    inputAudioFormat: realtimeAudioFormatSchema,
+    outputAudioFormat: realtimeAudioFormatSchema.optional(),
+    turnDetection: realtimeTurnDetectionSchema,
+    inputAudioTranscription: realtimeInputTranscriptionSchema.optional(),
+    translation: realtimeTranslationSchema.optional(),
+    tools: z.array(toolDefinitionSchema).max(128).optional(),
+    toolChoice: toolChoiceSchema.optional(),
+    temperature: z.number().min(0).max(2).optional(),
+    /** Per response, not per session: the session's ceiling is `limits`. */
+    maxOutputTokens: z.number().int().positive().safe().optional(),
+})
+    .strict();
+/**
+ * What `session.update` may change mid-session.
+ *
+ * The session's kind, its audio formats and its voice are fixed at open —
+ * providers refuse to change a voice once it has spoken, and a format change
+ * mid-stream would make every buffered frame ambiguous — so they are not fields
+ * here, and the strict parse refuses them rather than ignoring them.
+ */
+export const realtimeSessionConfigUpdateSchema = z
+    .object({
+    instructions: z.string().max(32_768).optional(),
+    outputModalities: z.array(realtimeOutputModalitySchema).min(1).max(2).optional(),
+    turnDetection: realtimeTurnDetectionSchema.optional(),
+    inputAudioTranscription: realtimeInputTranscriptionSchema.optional(),
+    tools: z.array(toolDefinitionSchema).max(128).optional(),
+    toolChoice: toolChoiceSchema.optional(),
+    temperature: z.number().min(0).max(2).optional(),
+    maxOutputTokens: z.number().int().positive().safe().optional(),
+})
+    .strict()
+    .superRefine((update, ctx) => {
+    if (Object.keys(update).length === 0) {
+        ctx.addIssue({
+            code: 'custom',
+            message: 'a session update changes at least one field',
+        });
+    }
+});
+/**
+ * The ceilings Oxy signed the session for. The data plane enforces each one and
+ * closes the session (`limit_exceeded`, `max_duration`, `idle_timeout`) rather
+ * than exceeding it: these bound what the session can cost, which is what Oxy's
+ * spend reservation was sized against.
+ */
+export const realtimeSessionLimitsSchema = z
+    .object({
+    maxDurationMs: z.number().int().min(1_000).max(MAX_REALTIME_SESSION_DURATION_MS),
+    idleTimeoutMs: z.number().int().min(1_000).max(MAX_REALTIME_SESSION_DURATION_MS),
+    maxInputAudioBytes: z.number().int().positive().max(MAX_REALTIME_SESSION_AUDIO_BYTES),
+    maxOutputAudioBytes: z.number().int().positive().max(MAX_REALTIME_SESSION_AUDIO_BYTES),
+    maxResponses: z.number().int().positive().max(10_000),
+})
+    .strict();
+/**
+ * One part of a conversation message.
+ *
+ * `input_audio` carries its audio inline (one bounded frame); `output_audio`
+ * never does — spoken output arrives as `output_audio.delta` events, and an
+ * item echoing it carries only the format and, once known, the transcript.
+ */
+export const realtimeContentPartSchema = z.discriminatedUnion('type', [
+    z.object({ type: z.literal('input_text'), text: z.string().max(1_048_576) }).strict(),
+    z
+        .object({
+        type: z.literal('input_audio'),
+        format: realtimeAudioFormatSchema,
+        data: audioFrameSchema.optional(),
+        transcript: z.string().optional(),
+    })
+        .strict(),
+    z.object({ type: z.literal('output_text'), text: z.string() }).strict(),
+    z
+        .object({
+        type: z.literal('output_audio'),
+        format: realtimeAudioFormatSchema,
+        transcript: z.string().optional(),
+    })
+        .strict(),
+]);
+const INPUT_PARTS = new Set(['input_text', 'input_audio']);
+/**
+ * One item of the conversation a session holds.
+ *
+ * `itemId` is optional on an item a client CREATES (the server assigns one) and
+ * is always stated by the server on the events that carry an item.
+ */
+export const realtimeConversationItemSchema = z
+    .discriminatedUnion('type', [
+    z
+        .object({
+        type: z.literal('message'),
+        itemId: realtimeItemIdSchema.optional(),
+        role: z.enum(['system', 'user', 'assistant']),
+        content: z.array(realtimeContentPartSchema).min(1).max(64),
+    })
+        .strict(),
+    z
+        .object({
+        type: z.literal('function_call'),
+        itemId: realtimeItemIdSchema.optional(),
+        callId: z.string().min(1).max(128),
+        name: z.string().min(1).max(128),
+        arguments: z.string(),
+    })
+        .strict(),
+    z
+        .object({
+        type: z.literal('function_call_output'),
+        itemId: realtimeItemIdSchema.optional(),
+        callId: z.string().min(1).max(128),
+        output: z.string().max(1_048_576),
+    })
+        .strict(),
+])
+    .superRefine((item, ctx) => {
+    if (item.type !== 'message')
+        return;
+    for (const [index, part] of item.content.entries()) {
+        const isInput = INPUT_PARTS.has(part.type);
+        // A user or system turn is input; an assistant turn is output. A part on
+        // the wrong side is one every provider would silently ignore.
+        if (item.role === 'assistant' ? isInput : !isInput) {
+            ctx.addIssue({
+                code: 'custom',
+                path: ['content', index, 'type'],
+                message: `a ${item.role} message cannot carry ${part.type}`,
+            });
+        }
+        if (item.role === 'system' && part.type !== 'input_text') {
+            ctx.addIssue({
+                code: 'custom',
+                path: ['content', index, 'type'],
+                message: 'a system message is text',
+            });
+        }
+    }
+});
+/** Per-response overrides for `response.create`. Absent fields keep the session's. */
+export const realtimeResponseParametersSchema = z
+    .object({
+    instructions: z.string().max(32_768).optional(),
+    outputModalities: z.array(realtimeOutputModalitySchema).min(1).max(2).optional(),
+    maxOutputTokens: z.number().int().positive().safe().optional(),
+    toolChoice: toolChoiceSchema.optional(),
+})
+    .strict();
+const commandBase = {
+    /** See `version.ts`: each command is a whole message on the wire. */
+    schemaVersion: z.literal(1),
+    requestId: requestIdSchema,
+    commandId: realtimeCommandIdSchema,
+};
+export const realtimeSessionUpdateCommandSchema = z.object({
+    ...commandBase,
+    type: z.literal('session.update'),
+    config: realtimeSessionConfigUpdateSchema,
+});
+export const realtimeItemCreateCommandSchema = z.object({
+    ...commandBase,
+    type: z.literal('conversation.item.create'),
+    /** Insert after this item; absent appends. */
+    previousItemId: realtimeItemIdSchema.optional(),
+    item: realtimeConversationItemSchema,
+});
+export const realtimeItemDeleteCommandSchema = z.object({
+    ...commandBase,
+    type: z.literal('conversation.item.delete'),
+    itemId: realtimeItemIdSchema,
+});
+/**
+ * Cut an assistant's audio item at the point the listener actually heard, after
+ * an interruption, so the conversation the model remembers is the one that was
+ * spoken rather than the one that was generated.
+ */
+export const realtimeItemTruncateCommandSchema = z.object({
+    ...commandBase,
+    type: z.literal('conversation.item.truncate'),
+    itemId: realtimeItemIdSchema,
+    contentIndex: z.number().int().nonnegative().safe(),
+    audioEndMs: milliseconds,
+});
+export const realtimeInputAudioAppendCommandSchema = z.object({
+    ...commandBase,
+    type: z.literal('input_audio.append'),
+    /** One bounded frame in the session's `inputAudioFormat`. */
+    data: audioFrameSchema,
+});
+export const realtimeInputAudioCommitCommandSchema = z.object({
+    ...commandBase,
+    type: z.literal('input_audio.commit'),
+});
+export const realtimeInputAudioClearCommandSchema = z.object({
+    ...commandBase,
+    type: z.literal('input_audio.clear'),
+});
+export const realtimeResponseCreateCommandSchema = z.object({
+    ...commandBase,
+    type: z.literal('response.create'),
+    response: realtimeResponseParametersSchema.optional(),
+});
+export const realtimeResponseCancelCommandSchema = z.object({
+    ...commandBase,
+    type: z.literal('response.cancel'),
+    /** Absent cancels the response in progress, if any. */
+    responseId: realtimeResponseIdSchema.optional(),
+});
+/**
+ * The first frame of a reconnection: replay every event after `afterSequence`.
+ * `-1` asks for every buffered event from the start.
+ */
+export const realtimeSessionResumeCommandSchema = z.object({
+    ...commandBase,
+    type: z.literal('session.resume'),
+    afterSequence: z.number().int().min(-1).safe(),
+});
+export const realtimeSessionCloseCommandSchema = z.object({
+    ...commandBase,
+    type: z.literal('session.close'),
+});
+/** Every command a client can send. */
+export const realtimeClientCommandSchema = z.discriminatedUnion('type', [
+    realtimeSessionUpdateCommandSchema,
+    realtimeItemCreateCommandSchema,
+    realtimeItemDeleteCommandSchema,
+    realtimeItemTruncateCommandSchema,
+    realtimeInputAudioAppendCommandSchema,
+    realtimeInputAudioCommitCommandSchema,
+    realtimeInputAudioClearCommandSchema,
+    realtimeResponseCreateCommandSchema,
+    realtimeResponseCancelCommandSchema,
+    realtimeSessionResumeCommandSchema,
+    realtimeSessionCloseCommandSchema,
+]);
+const eventBase = {
+    /** See `version.ts`: each event is a whole message on the wire. */
+    schemaVersion: z.literal(1),
+    requestId: requestIdSchema,
+    /** One monotonic sequence per session, from 0; what makes a replay detectable. */
+    sequence: sequenceSchema,
+    /** The command this event answers, when it answers one. */
+    commandId: realtimeCommandIdSchema.optional(),
+};
+export const realtimeSessionUpdatedEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('session.updated'),
+    config: realtimeSessionConfigSchema,
+});
+export const realtimeSessionResumedEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('session.resumed'),
+    afterSequence: z.number().int().min(-1).safe(),
+});
+/** Sent before a command is applied. `duplicate` means it was not applied again. */
+export const realtimeCommandAcceptedEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('command.accepted'),
+    commandId: realtimeCommandIdSchema,
+    duplicate: z.boolean(),
+});
+export const realtimeItemAddedEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('conversation.item.added'),
+    itemId: realtimeItemIdSchema,
+    previousItemId: realtimeItemIdSchema.optional(),
+    item: realtimeConversationItemSchema,
+});
+export const realtimeItemDoneEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('conversation.item.done'),
+    itemId: realtimeItemIdSchema,
+    item: realtimeConversationItemSchema,
+});
+export const realtimeItemDeletedEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('conversation.item.deleted'),
+    itemId: realtimeItemIdSchema,
+});
+export const realtimeItemTruncatedEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('conversation.item.truncated'),
+    itemId: realtimeItemIdSchema,
+    contentIndex: z.number().int().nonnegative().safe(),
+    audioEndMs: milliseconds,
+});
+/**
+ * Voice activity began. With `interruptResponse`, a response in progress is
+ * cancelled by it: this event is the interruption, and the client should stop
+ * playback and truncate what was not heard.
+ */
+export const realtimeSpeechStartedEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('input_audio.speech_started'),
+    itemId: realtimeItemIdSchema,
+    audioStartMs: milliseconds,
+});
+export const realtimeSpeechStoppedEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('input_audio.speech_stopped'),
+    itemId: realtimeItemIdSchema,
+    audioEndMs: milliseconds,
+});
+export const realtimeInputAudioCommittedEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('input_audio.committed'),
+    itemId: realtimeItemIdSchema,
+    previousItemId: realtimeItemIdSchema.optional(),
+});
+export const realtimeInputAudioClearedEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('input_audio.cleared'),
+});
+export const realtimeResponseCreatedEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('response.created'),
+    responseId: realtimeResponseIdSchema,
+});
+/** One bounded frame of spoken output, in the session's `outputAudioFormat`. */
+export const realtimeOutputAudioDeltaEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('output_audio.delta'),
+    responseId: realtimeResponseIdSchema,
+    itemId: realtimeItemIdSchema,
+    contentIndex: z.number().int().nonnegative().safe(),
+    format: realtimeAudioFormatSchema,
+    data: audioFrameSchema,
+});
+export const realtimeOutputAudioDoneEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('output_audio.done'),
+    responseId: realtimeResponseIdSchema,
+    itemId: realtimeItemIdSchema,
+    contentIndex: z.number().int().nonnegative().safe(),
+});
+/**
+ * Which audio a transcript describes: the caller's (`input_audio`) or the
+ * model's own speech (`output_audio`). Neither is answer text.
+ */
+export const realtimeTranscriptSourceSchema = z.enum(['input_audio', 'output_audio']);
+export const realtimeTranscriptDeltaEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('transcript.delta'),
+    source: realtimeTranscriptSourceSchema,
+    itemId: realtimeItemIdSchema,
+    contentIndex: z.number().int().nonnegative().safe(),
+    /** Present for `output_audio`: the response whose speech this is. */
+    responseId: realtimeResponseIdSchema.optional(),
+    text: z.string(),
+});
+export const realtimeTranscriptDoneEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('transcript.done'),
+    source: realtimeTranscriptSourceSchema,
+    itemId: realtimeItemIdSchema,
+    contentIndex: z.number().int().nonnegative().safe(),
+    responseId: realtimeResponseIdSchema.optional(),
+    transcript: z.string(),
+});
+/** Written output of a response (`outputModalities` including `text`). */
+export const realtimeTextDeltaEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('text.delta'),
+    responseId: realtimeResponseIdSchema,
+    itemId: realtimeItemIdSchema,
+    contentIndex: z.number().int().nonnegative().safe(),
+    text: z.string(),
+});
+/** A tool call being streamed; same accumulation rules as the one-shot event. */
+export const realtimeToolCallEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('tool_call'),
+    responseId: realtimeResponseIdSchema,
+    itemId: realtimeItemIdSchema,
+    toolCallId: z.string().min(1).max(128),
+    name: z.string().min(1).max(128).optional(),
+    argumentsDelta: z.string().optional(),
+    complete: z.boolean(),
+});
+/** How a response ended. */
+export const realtimeResponseStatusSchema = z.enum([
+    'completed',
+    'cancelled',
+    'incomplete',
+    'failed',
+]);
+/** Why a session ended. */
+export const realtimeSessionCloseReasonSchema = z.enum([
+    'client_closed',
+    'max_duration',
+    'idle_timeout',
+    'limit_exceeded',
+    'resume_expired',
+    'upstream_closed',
+    'upstream_error',
+    'no_route_available',
+    'server_shutdown',
+]);
+/**
+ * What a session kind requires of its configuration. Each kind needs a
+ * different subset of the config, so these rules are cross-field and live
+ * here once: the public `session.open` frame and the signed session request
+ * apply the same ones, and a session a client may open is one Kaana may run.
+ */
+export function realtimeConfigIssues(kind, config) {
+    const issues = [];
+    const found = (path, message) => issues.push({ path, message });
+    const producesAudio = config.outputModalities?.includes('audio') === true;
+    const hasTools = config.tools !== undefined && config.tools.length > 0;
+    switch (kind) {
+        case 'conversation':
+            if (config.outputModalities === undefined) {
+                found(['outputModalities'], 'a conversation names what its responses produce');
+            }
+            if (producesAudio && (config.voice === undefined || config.outputAudioFormat === undefined)) {
+                found([], 'spoken responses need a voice and an output audio format');
+            }
+            if (config.translation !== undefined) {
+                found(['translation'], 'only a translation session translates');
+            }
+            break;
+        case 'transcription':
+            if (config.inputAudioTranscription === undefined) {
+                found(['inputAudioTranscription'], 'a transcription session transcribes its input');
+            }
+            for (const field of [
+                'outputModalities',
+                'voice',
+                'outputAudioFormat',
+                'translation',
+                'tools',
+                'toolChoice',
+                'maxOutputTokens',
+            ]) {
+                if (config[field] !== undefined) {
+                    found([field], 'a transcription session never responds');
+                }
+            }
+            if (config.turnDetection.type !== 'none' &&
+                (config.turnDetection.createResponse || config.turnDetection.interruptResponse)) {
+                found(['turnDetection'], 'a transcription session has no response to create or interrupt');
+            }
+            break;
+        case 'translation':
+            if (config.translation === undefined || config.outputAudioFormat === undefined) {
+                found([], 'a translation session names its target language and output audio format');
+            }
+            if (config.tools !== undefined || config.toolChoice !== undefined) {
+                found(['tools'], 'a translation session calls no tools');
+            }
+            break;
+    }
+    if (config.toolChoice !== undefined && !hasTools) {
+        found(['toolChoice'], 'a tool choice requires at least one tool definition');
+    }
+    if (config.tools !== undefined) {
+        const names = config.tools.map((tool) => tool.name);
+        if (new Set(names).size !== names.length) {
+            found(['tools'], 'tool names must be unique within one session');
+        }
+    }
+    if (config.outputModalities !== undefined &&
+        new Set(config.outputModalities).size !== config.outputModalities.length) {
+        found(['outputModalities'], 'each output modality is named at most once');
+    }
+    return issues;
+}
+/* -------------------------------------------------------------------------- */
+/*  Opening a session (`GET /v1/realtime`)                                     */
+/* -------------------------------------------------------------------------- */
+/**
+ * The most text `conversation.item.create` commands a session may be opened
+ * for: the bound a duration-priced route's hold is sized from.
+ */
+export const MAX_REALTIME_TEXT_ITEMS = 10_000;
+/**
+ * The customer's partial limits request: every signed limit, each filled from
+ * Kaana's defaults when omitted, plus `maxTextItems`, the most billable text
+ * items the session may send (absent, it defaults to `maxResponses`).
+ */
+export const realtimeOpenLimitsSchema = realtimeSessionLimitsSchema
+    .partial()
+    .extend({
+    maxTextItems: z.number().int().nonnegative().max(MAX_REALTIME_TEXT_ITEMS).optional(),
+})
+    .strict();
+/**
+ * The first frame a client sends on `GET /v1/realtime?model=…` to OPEN a
+ * session (a reconnection instead sends `session.resume`). One grant covers the
+ * whole session and it is settled exactly once, when the session closes.
+ *
+ * Moved from the Oxy edge (`realtimeOpenFrameSchema`) by Kaana ADR 0002.
+ */
+export const realtimeSessionOpenSchema = z
+    .object({
+    type: z.literal('session.open'),
+    kind: realtimeSessionKindSchema,
+    /** Absent means `websocket`, the only transport there is. */
+    transport: realtimeSessionTransportSchema.optional(),
+    config: realtimeSessionConfigSchema,
+    limits: realtimeOpenLimitsSchema.optional(),
+    /** The caller's own correlation id. */
+    clientSessionId: z.string().min(1).max(128).optional(),
+    labels: requestLabelsSchema.optional(),
+})
+    .strict()
+    .superRefine((open, ctx) => {
+    for (const found of realtimeConfigIssues(open.kind, open.config)) {
+        ctx.addIssue({ code: 'custom', path: ['config', ...found.path], message: found.message });
+    }
+    const limits = open.limits;
+    if (limits?.idleTimeoutMs !== undefined &&
+        limits.maxDurationMs !== undefined &&
+        limits.idleTimeoutMs > limits.maxDurationMs) {
+        ctx.addIssue({
+            code: 'custom',
+            path: ['limits', 'idleTimeoutMs'],
+            message: 'an idle timeout cannot outlast the session',
+        });
+    }
+});
+/**
+ * Check a `session.open` against the model's published catalogue entry — its
+ * `capabilities.realtime` and its `audio` — the realtime counterpart of
+ * {@link checkSpeechRequest} (Kaana ADR 0002 decision 9; Alia ADR 0017 §5).
+ * Kaana refuses an open with `invalid_request` and the first issue's `param`
+ * BEFORE asking Oxy for a grant; a client may run the same check to choose a
+ * model.
+ *
+ * A field the catalogue leaves undeclared (an older producer) is not checked:
+ * absent is unknown, never "everything" and never "nothing". The one
+ * exception is input transcription in a conversation, which a model must
+ * declare to be asked for, because asking for it starts a charge.
+ *
+ * Returns every issue, in field order; an empty list means the model accepts
+ * the open.
+ */
+export function checkRealtimeSessionOpen(open, model) {
+    const realtime = model.capabilities.realtime;
+    if (realtime === undefined) {
+        return [{ param: 'kind', message: 'this model holds no realtime session' }];
+    }
+    const issues = [];
+    if (!realtime.sessionKinds.includes(open.kind)) {
+        issues.push({ param: 'kind', message: `this model holds no ${open.kind} session` });
+    }
+    const transport = open.transport ?? 'websocket';
+    if (!realtime.transports.includes(transport)) {
+        issues.push({ param: 'transport', message: `this model is not served over ${transport}` });
+    }
+    const { config } = open;
+    const formats = realtime.audioFormats;
+    if (formats !== undefined && !formats.input.includes(config.inputAudioFormat)) {
+        issues.push({
+            param: 'config.inputAudioFormat',
+            message: `this model does not take ${config.inputAudioFormat} input`,
+        });
+    }
+    if (formats !== undefined &&
+        config.outputAudioFormat !== undefined &&
+        !formats.output.includes(config.outputAudioFormat)) {
+        issues.push({
+            param: 'config.outputAudioFormat',
+            message: `this model does not produce ${config.outputAudioFormat}`,
+        });
+    }
+    const audio = model.audio;
+    if (audio !== undefined &&
+        config.voice !== undefined &&
+        !audio.voices.some((voice) => voice.id === config.voice)) {
+        issues.push({ param: 'config.voice', message: `this model has no voice ${config.voice}` });
+    }
+    const transcription = config.inputAudioTranscription;
+    if (transcription !== undefined &&
+        open.kind === 'conversation' &&
+        realtime.inputAudioTranscription !== true) {
+        issues.push({
+            param: 'config.inputAudioTranscription',
+            message: 'this model does not transcribe its input inside a conversation',
+        });
+    }
+    if (transcription?.language !== undefined &&
+        audio !== undefined &&
+        !audio.languages.includes(transcription.language)) {
+        issues.push({
+            param: 'config.inputAudioTranscription.language',
+            message: `this model does not declare the language ${transcription.language}`,
+        });
+    }
+    return issues;
+}
+/* -------------------------------------------------------------------------- */
+/*  Server events a public client reads                                        */
+/* -------------------------------------------------------------------------- */
+/**
+ * The session opened: the revision-pinned model, fixed for the rest of the
+ * session, and the effective configuration and limits. The signed lane's
+ * `realtimeSessionCreatedEventSchema` also names the provider and deployment;
+ * a public client is never told either.
+ */
+export const realtimePublicSessionCreatedEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('session.created'),
+    /** The revision serving the session. */
+    model: modelReferenceSchema,
+    kind: realtimeSessionKindSchema,
+    config: realtimeSessionConfigSchema,
+    limits: realtimeSessionLimitsSchema,
+    resumeWindowMs: z.number().int().nonnegative().max(MAX_REALTIME_RESUME_WINDOW_MS),
+    startedAt: inferenceTimestampSchema,
+    expiresAt: inferenceTimestampSchema,
+});
+/** A response ended, with the units it consumed. */
+export const realtimePublicResponseDoneEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('response.done'),
+    responseId: realtimeResponseIdSchema,
+    status: realtimeResponseStatusSchema,
+    finishReason: inferenceFinishReasonSchema.optional(),
+    units: z.array(usageQuantitySchema),
+    usageSource: usageSourceSchema,
+});
+/**
+ * An error in the public vocabulary. `fatal: false` refuses one command and
+ * the session continues; `fatal: true` is followed by `session.closed`.
+ */
+export const realtimePublicErrorEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('error'),
+    fatal: z.boolean(),
+    error: publicErrorSchema,
+});
+/** The terminal event, with the session's total units. */
+export const realtimePublicSessionClosedEventSchema = z.object({
+    ...eventBase,
+    type: z.literal('session.closed'),
+    reason: realtimeSessionCloseReasonSchema,
+    units: z.array(usageQuantitySchema),
+    usageSource: usageSourceSchema,
+    closedAt: inferenceTimestampSchema,
+});
+/** Every event a public session can emit. */
+export const realtimePublicServerEventSchema = z
+    .discriminatedUnion('type', [
+    realtimePublicSessionCreatedEventSchema,
+    realtimeSessionUpdatedEventSchema,
+    realtimeSessionResumedEventSchema,
+    realtimeCommandAcceptedEventSchema,
+    realtimeItemAddedEventSchema,
+    realtimeItemDoneEventSchema,
+    realtimeItemDeletedEventSchema,
+    realtimeItemTruncatedEventSchema,
+    realtimeSpeechStartedEventSchema,
+    realtimeSpeechStoppedEventSchema,
+    realtimeInputAudioCommittedEventSchema,
+    realtimeInputAudioClearedEventSchema,
+    realtimeResponseCreatedEventSchema,
+    realtimeOutputAudioDeltaEventSchema,
+    realtimeOutputAudioDoneEventSchema,
+    realtimeTranscriptDeltaEventSchema,
+    realtimeTranscriptDoneEventSchema,
+    realtimeTextDeltaEventSchema,
+    realtimeToolCallEventSchema,
+    realtimePublicResponseDoneEventSchema,
+    realtimePublicErrorEventSchema,
+    realtimePublicSessionClosedEventSchema,
+])
+    .superRefine((event, ctx) => {
+    if (event.type !== 'response.done' && event.type !== 'session.closed')
+        return;
+    const units = event.units.map((quantity) => quantity.unit);
+    if (new Set(units).size !== units.length) {
+        ctx.addIssue({
+            code: 'custom',
+            path: ['units'],
+            message: 'each unit is reported at most once',
+        });
+    }
+});