@kaana.ai/contracts 4.4.1 → 6.0.0

package/dist/internal/catalogue.d.ts

@@ -783,6 +783,10 @@ export declare const declaredModelSchema: z.ZodObject<{
             currency: z.ZodString;
         }, z.core.$strict>>;
     }, z.core.$strict>>;
+    pricePerMinute: z.ZodOptional<z.ZodObject<{
+        amount: z.core.$ZodBranded<z.ZodString, "ExactDecimal", "out">;
+        currency: z.ZodString;
+    }, z.core.$strict>>;
     evaluations: z.ZodDefault<z.ZodArray<z.ZodObject<{
         suite: z.ZodString;
         metric: z.ZodString;
@@ -1007,6 +1011,10 @@ export declare const declaredCatalogueSchema: z.ZodObject<{
                 currency: z.ZodString;
             }, z.core.$strict>>;
         }, z.core.$strict>>;
+        pricePerMinute: z.ZodOptional<z.ZodObject<{
+            amount: z.core.$ZodBranded<z.ZodString, "ExactDecimal", "out">;
+            currency: z.ZodString;
+        }, z.core.$strict>>;
         evaluations: z.ZodDefault<z.ZodArray<z.ZodObject<{
             suite: z.ZodString;
             metric: z.ZodString;

package/dist/internal/realtime.d.ts

@@ -104,7 +104,6 @@ export declare const realtimeSessionRequestSchema: z.ZodObject<{
             name: z.ZodString;
             description: z.ZodOptional<z.ZodString>;
             parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
-            strict: z.ZodOptional<z.ZodBoolean>;
         }, z.core.$strict>>>;
         toolChoice: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
             auto: "auto";
@@ -114,7 +113,6 @@ export declare const realtimeSessionRequestSchema: z.ZodObject<{
             type: z.ZodLiteral<"function">;
             name: z.ZodString;
         }, z.core.$strict>]>>;
-        temperature: z.ZodOptional<z.ZodNumber>;
         maxOutputTokens: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>;
     limits: z.ZodObject<{
@@ -246,7 +244,6 @@ export declare const realtimeSessionCreatedEventSchema: z.ZodObject<{
             name: z.ZodString;
             description: z.ZodOptional<z.ZodString>;
             parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
-            strict: z.ZodOptional<z.ZodBoolean>;
         }, z.core.$strict>>>;
         toolChoice: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
             auto: "auto";
@@ -256,7 +253,6 @@ export declare const realtimeSessionCreatedEventSchema: z.ZodObject<{
             type: z.ZodLiteral<"function">;
             name: z.ZodString;
         }, z.core.$strict>]>>;
-        temperature: z.ZodOptional<z.ZodNumber>;
         maxOutputTokens: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>;
     limits: z.ZodObject<{
@@ -505,7 +501,6 @@ export declare const realtimeServerEventSchema: z.ZodDiscriminatedUnion<[z.ZodOb
             name: z.ZodString;
             description: z.ZodOptional<z.ZodString>;
             parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
-            strict: z.ZodOptional<z.ZodBoolean>;
         }, z.core.$strict>>>;
         toolChoice: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
             auto: "auto";
@@ -515,7 +510,6 @@ export declare const realtimeServerEventSchema: z.ZodDiscriminatedUnion<[z.ZodOb
             type: z.ZodLiteral<"function">;
             name: z.ZodString;
         }, z.core.$strict>]>>;
-        temperature: z.ZodOptional<z.ZodNumber>;
         maxOutputTokens: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>;
     limits: z.ZodObject<{
@@ -583,7 +577,6 @@ export declare const realtimeServerEventSchema: z.ZodDiscriminatedUnion<[z.ZodOb
             name: z.ZodString;
             description: z.ZodOptional<z.ZodString>;
             parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
-            strict: z.ZodOptional<z.ZodBoolean>;
         }, z.core.$strict>>>;
         toolChoice: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
             auto: "auto";
@@ -593,7 +586,6 @@ export declare const realtimeServerEventSchema: z.ZodDiscriminatedUnion<[z.ZodOb
             type: z.ZodLiteral<"function">;
             name: z.ZodString;
         }, z.core.$strict>]>>;
-        temperature: z.ZodOptional<z.ZodNumber>;
         maxOutputTokens: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>;
     schemaVersion: z.ZodLiteral<1>;

package/dist/internal/request.d.ts

@@ -344,6 +344,10 @@ export declare const inferenceRequestSchema: z.ZodObject<{
         }>;
         speed: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>>;
+    transcription: z.ZodOptional<z.ZodObject<{
+        language: z.ZodOptional<z.ZodString>;
+        prompt: z.ZodOptional<z.ZodString>;
+    }, z.core.$strict>>;
     audioOutput: z.ZodOptional<z.ZodObject<{
         voice: z.ZodString;
         format: z.ZodEnum<{
@@ -619,6 +623,10 @@ export declare const scopedInferenceRequestSchema: z.ZodObject<{
         }>;
         speed: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>>;
+    transcription: z.ZodOptional<z.ZodObject<{
+        language: z.ZodOptional<z.ZodString>;
+        prompt: z.ZodOptional<z.ZodString>;
+    }, z.core.$strict>>;
     audioOutput: z.ZodOptional<z.ZodObject<{
         voice: z.ZodString;
         format: z.ZodEnum<{
@@ -864,6 +872,10 @@ export declare const privateAutoInferenceRequestSchema: z.ZodObject<{
             high: "high";
         }>;
     }, z.core.$strict>>;
+    transcription: z.ZodOptional<z.ZodObject<{
+        language: z.ZodOptional<z.ZodString>;
+        prompt: z.ZodOptional<z.ZodString>;
+    }, z.core.$strict>>;
     tools: z.ZodDefault<z.ZodArray<z.ZodObject<{
         type: z.ZodLiteral<"function">;
         name: z.ZodString;

package/dist/internal/request.js

@@ -7,7 +7,7 @@ import { z } from 'zod';
 import { inferenceApiFormatSchema, inferenceModalitySchema } from '../public/catalogue.js';
 import { decisionInputSchema } from '../public/decisions.js';
 import { idempotencyKeySchema, inferenceTimestampSchema } from '../public/identifiers.js';
-import { inferenceAudioOutputParametersSchema, inferenceMessageSchema, inferenceReasoningSchema, inferenceSpeechParametersSchema, responseFormatSchema, samplingParametersSchema, toolChoiceSchema, toolDefinitionSchema, } from '../public/messages.js';
+import { inferenceAudioOutputParametersSchema, inferenceMessageSchema, inferenceReasoningSchema, inferenceSpeechParametersSchema, inferenceTranscriptionParametersSchema, responseFormatSchema, samplingParametersSchema, toolChoiceSchema, toolDefinitionSchema, } from '../public/messages.js';
 import { routingPolicyReferenceSchema } from '../public/policy.js';
 import { inferenceAttributionSchema } from './attribution.js';
 import { PRIVATE_AUTO_REQUEST_ENVELOPE_VERSION, privateAutoExecutionSchema, privateAutoInputSchema, } from './privateAutoExecution.js';
@@ -96,6 +96,8 @@ const inferenceRequestObjectSchema = z.object({
     /** Absent means the route's own default reasoning behaviour. */
     reasoning: inferenceReasoningSchema.optional(),
     speech: inferenceSpeechParametersSchema.optional(),
+    /** File transcription parameters. Contract set 4.5.0. */
+    transcription: inferenceTranscriptionParametersSchema.optional(),
     /** Spoken output from a conversational model. Contract set 3.2.0. */
     audioOutput: inferenceAudioOutputParametersSchema.optional(),
     tools: z.array(toolDefinitionSchema).default([]),
@@ -164,6 +166,7 @@ export const inferenceRequestSchema = inferenceRequestObjectSchema.superRefine((
             request.responseFormat !== undefined ||
             request.reasoning !== undefined ||
             request.speech !== undefined ||
+            request.transcription !== undefined ||
             request.audioOutput !== undefined) {
             ctx.addIssue({
                 code: 'custom',
@@ -191,6 +194,16 @@ export const inferenceRequestSchema = inferenceRequestObjectSchema.superRefine((
             message: 'speech parameters require the audio_speech API format',
         });
     }
+    if (request.transcription !== undefined &&
+        (request.client.apiFormat !== 'audio_transcriptions' ||
+            request.modality !== 'audio' ||
+            request.input.format !== 'messages')) {
+        ctx.addIssue({
+            code: 'custom',
+            path: ['transcription'],
+            message: 'transcription parameters require the audio_transcriptions API format with audio messages input',
+        });
+    }
     if (request.audioOutput !== undefined) {
         if (request.client.apiFormat !== 'chat_completions') {
             ctx.addIssue({

package/dist/public/audio.d.ts

@@ -8,6 +8,7 @@
  * (decision 9) and transcription are new in contract set 4.0.0.
  */
 import { z } from 'zod';
+import { type AudioInputMediaType } from './audioFormats.js';
 import type { ModelAudioCapabilities } from './catalogue.js';
 /** The longest text one speech request may carry, before any model's own limit. */
 export declare const MAX_SPEECH_INPUT_CHARACTERS = 100000;
@@ -248,11 +249,6 @@ export declare const transcriptionRequestSchema: z.ZodObject<{
     model: z.ZodString;
     language: z.ZodOptional<z.ZodString>;
     prompt: z.ZodOptional<z.ZodString>;
-    timestamps: z.ZodOptional<z.ZodEnum<{
-        none: "none";
-        segment: "segment";
-        word: "word";
-    }>>;
 }, z.core.$strict>;
 /**
  * The longest audio one transcription request may carry, when its container
@@ -277,11 +273,6 @@ export declare const transcriptionJsonRequestSchema: z.ZodObject<{
     model: z.ZodString;
     language: z.ZodOptional<z.ZodString>;
     prompt: z.ZodOptional<z.ZodString>;
-    timestamps: z.ZodOptional<z.ZodEnum<{
-        none: "none";
-        segment: "segment";
-        word: "word";
-    }>>;
     audio: z.ZodObject<{
         mediaType: z.ZodEnum<{
             "audio/mpeg": "audio/mpeg";
@@ -296,7 +287,7 @@ export declare const transcriptionJsonRequestSchema: z.ZodObject<{
         data: z.ZodString;
     }, z.core.$strict>;
 }, z.core.$strict>;
-/** The response. Segments or words are present only when the request asked for them. */
+/** The response. */
 export declare const transcriptionSchema: z.ZodObject<{
     schemaVersion: z.ZodLiteral<1>;
     requestId: z.ZodString;
@@ -304,16 +295,6 @@ export declare const transcriptionSchema: z.ZodObject<{
     text: z.ZodString;
     language: z.ZodOptional<z.ZodString>;
     durationMs: z.ZodOptional<z.ZodNumber>;
-    segments: z.ZodOptional<z.ZodArray<z.ZodObject<{
-        text: z.ZodString;
-        startMs: z.ZodNumber;
-        endMs: z.ZodNumber;
-    }, z.core.$strict>>>;
-    words: z.ZodOptional<z.ZodArray<z.ZodObject<{
-        word: z.ZodString;
-        startMs: z.ZodNumber;
-        endMs: z.ZodNumber;
-    }, z.core.$strict>>>;
     usage: z.ZodArray<z.ZodObject<{
         unit: z.ZodEnum<{
             input_tokens: "input_tokens";
@@ -336,6 +317,22 @@ export declare const transcriptionSchema: z.ZodObject<{
         quantity: z.ZodNumber;
     }, z.core.$strict>>;
 }, z.core.$strict>;
+/** Why a transcription request cannot be served by a model, before anything is held. */
+export interface TranscriptionRequestIssue {
+    /** The request's field at fault, as `invalid_request.param` names it. */
+    readonly param: 'file';
+    readonly message: string;
+}
+/**
+ * Check a transcription upload against the model's published `audio` — the
+ * transcription counterpart of {@link checkSpeechRequest}. Kaana refuses an
+ * upload in a format the model does not read with `invalid_request` and
+ * `param: 'file'` BEFORE asking Oxy for a grant. A language or a prompt is
+ * never refused: Kaana sends the provider its spelling of the language, or
+ * none for the model to detect, and the prompt where it is taken (Kaana ADR
+ * 0003). `mediaType` is the upload's.
+ */
+export declare function checkTranscriptionRequest(mediaType: AudioInputMediaType, audio: ModelAudioCapabilities): TranscriptionRequestIssue[];
 export type SpeechStreamFormat = z.infer<typeof speechStreamFormatSchema>;
 export type SpeechRequest = z.infer<typeof speechRequestSchema>;
 export type SpeechAudioDeltaEvent = z.infer<typeof speechAudioDeltaEventSchema>;

package/dist/public/audio.js

@@ -160,12 +160,16 @@ export const MAX_TRANSCRIPTION_FILE_BYTES = 20 * 1024 * 1024;
 export const transcriptionRequestSchema = z
     .object({
     model: modelReferenceSchema,
-    /** The spoken language, when known; validated against the model's `audio.languages`. */
+    /**
+     * The spoken language, when known, as the client's own BCP 47 tag. Kaana
+     * sends the provider its spelling of it, or none for the model to detect.
+     */
     language: languageTagSchema.optional(),
-    /** Vocabulary or context that helps the model; never an instruction. */
+    /**
+     * A hint — vocabulary or context that helps the model, never an
+     * instruction — used where the model takes one.
+     */
     prompt: z.string().max(4096).optional(),
-    /** Absent means `none`. */
-    timestamps: z.enum(['none', 'segment', 'word']).optional(),
 })
     .strict();
 /**
@@ -201,11 +205,7 @@ export const transcriptionJsonRequestSchema = transcriptionRequestSchema
         .strict(),
 })
     .strict();
-const timedSpanFields = {
-    startMs: z.number().int().nonnegative().safe(),
-    endMs: z.number().int().nonnegative().safe(),
-};
-/** The response. Segments or words are present only when the request asked for them. */
+/** The response. */
 export const transcriptionSchema = z
     .object({
     schemaVersion: z.literal(1),
@@ -213,24 +213,24 @@ export const transcriptionSchema = z
     /** The revision that actually served the request. */
     model: modelReferenceSchema,
     text: z.string(),
-    /** The language the model heard, when it reports one. */
+    /** The language the model heard, when its provider reports it as a tag. */
     language: languageTagSchema.optional(),
+    /** The audio's duration, measured from its container before the grant. */
     durationMs: z.number().int().nonnegative().safe().optional(),
-    segments: z.array(z.object({ ...timedSpanFields, text: z.string() }).strict()).optional(),
-    words: z.array(z.object({ ...timedSpanFields, word: z.string() }).strict()).optional(),
     usage: usageSchema,
 })
-    .strict()
-    .superRefine((transcription, ctx) => {
-    for (const field of ['segments', 'words']) {
-        for (const [index, span] of (transcription[field] ?? []).entries()) {
-            if (span.endMs < span.startMs) {
-                ctx.addIssue({
-                    code: 'custom',
-                    path: [field, index, 'endMs'],
-                    message: 'a span cannot end before it starts',
-                });
-            }
-        }
-    }
-});
+    .strict();
+/**
+ * Check a transcription upload against the model's published `audio` — the
+ * transcription counterpart of {@link checkSpeechRequest}. Kaana refuses an
+ * upload in a format the model does not read with `invalid_request` and
+ * `param: 'file'` BEFORE asking Oxy for a grant. A language or a prompt is
+ * never refused: Kaana sends the provider its spelling of the language, or
+ * none for the model to detect, and the prompt where it is taken (Kaana ADR
+ * 0003). `mediaType` is the upload's.
+ */
+export function checkTranscriptionRequest(mediaType, audio) {
+    return audio.audioFormats.input.includes(mediaType)
+        ? []
+        : [{ param: 'file', message: `this model does not take ${mediaType} audio` }];
+}

package/dist/public/audioFormats.d.ts

@@ -50,6 +50,14 @@ export declare const realtimeAudioFormatSchema: z.ZodEnum<{
     g711_ulaw: "g711_ulaw";
     g711_alaw: "g711_alaw";
 }>;
+/**
+ * What a session's responses produce. Declared here, beside the encodings,
+ * since contract set 5.0.0.
+ */
+export declare const realtimeOutputModalitySchema: z.ZodEnum<{
+    text: "text";
+    audio: "audio";
+}>;
 /**
  * The media type a speech request's `response_format` produces. A client
  * knows it before the first byte, which is why the speech stream events do

package/dist/public/audioFormats.js

@@ -42,6 +42,11 @@ export const audioInputMediaTypeSchema = z.enum([
  * catalogue cannot import the realtime module that imports it.
  */
 export const realtimeAudioFormatSchema = z.enum(['pcm16_24khz', 'g711_ulaw', 'g711_alaw']);
+/**
+ * What a session's responses produce. Declared here, beside the encodings,
+ * since contract set 5.0.0.
+ */
+export const realtimeOutputModalitySchema = z.enum(['text', 'audio']);
 const SPEECH_MEDIA_TYPES = {
     mp3: 'audio/mpeg',
     opus: 'audio/ogg',

package/dist/public/catalogue.d.ts

@@ -352,6 +352,12 @@ export declare const voiceSchema: z.ZodObject<{
  * none; a caller that needs a language then cannot choose the model.
  * `gender` is published only where it has been reviewed against the
  * provider's own catalogue; absent is unknown.
+ *
+ * A model's `languages` describe it; they are not a list a request must pick
+ * from. A request names the client's own BCP 47 tag, and Kaana sends the
+ * provider its spelling of that language, or none for the model to detect
+ * (Kaana ADR 0003). A transcription model's `audioFormats.input` are the
+ * uploads it reads.
  */
 export declare const modelAudioCapabilitiesSchema: z.ZodObject<{
     voices: z.ZodArray<z.ZodObject<{
@@ -573,6 +579,10 @@ export declare const modelSchema: z.ZodObject<{
             currency: z.ZodString;
         }, z.core.$strict>>;
     }, z.core.$strict>>;
+    pricePerMinute: z.ZodOptional<z.ZodObject<{
+        amount: z.core.$ZodBranded<z.ZodString, "ExactDecimal", "out">;
+        currency: z.ZodString;
+    }, z.core.$strict>>;
     deprecation: z.ZodObject<{
         status: z.ZodEnum<{
             active: "active";
@@ -788,6 +798,10 @@ export declare const modelListSchema: z.ZodObject<{
                 currency: z.ZodString;
             }, z.core.$strict>>;
         }, z.core.$strict>>;
+        pricePerMinute: z.ZodOptional<z.ZodObject<{
+            amount: z.core.$ZodBranded<z.ZodString, "ExactDecimal", "out">;
+            currency: z.ZodString;
+        }, z.core.$strict>>;
         deprecation: z.ZodObject<{
             status: z.ZodEnum<{
                 active: "active";

package/dist/public/catalogue.js

@@ -6,7 +6,7 @@
 import { z } from 'zod';
 import { inferenceDateSchema, inferenceHttpsUrlSchema, inferenceRegionSchema, inferenceTimestampSchema, modelIdSchema, modelReferenceSchema, modelRevisionLabelSchema, publisherSlugSchema, } from './identifiers.js';
 import { audioInputMediaTypeSchema, languageTagSchema, realtimeAudioFormatSchema, } from './audioFormats.js';
-import { priceSnapshotSchema } from './money.js';
+import { moneySchema, priceSnapshotSchema } from './money.js';
 import { inferenceAudioMediaTypeSchema } from './streamEvents.js';
 /** Input/request modalities; structured decisions are an output, never an input. */
 export const inferenceModalitySchema = z.enum(['text', 'image', 'audio', 'video', 'embedding']);
@@ -378,6 +378,12 @@ export const voiceSchema = z
  * none; a caller that needs a language then cannot choose the model.
  * `gender` is published only where it has been reviewed against the
  * provider's own catalogue; absent is unknown.
+ *
+ * A model's `languages` describe it; they are not a list a request must pick
+ * from. A request names the client's own BCP 47 tag, and Kaana sends the
+ * provider its spelling of that language, or none for the model to detect
+ * (Kaana ADR 0003). A transcription model's `audioFormats.input` are the
+ * uploads it reads.
  */
 export const modelAudioCapabilitiesSchema = z
     .object({
@@ -469,6 +475,14 @@ export const modelSchema = z.object({
     dataPolicy: inferenceDataPolicySchema,
     /** Oxy's list price. Absent for a model a customer cannot buy per unit (BYOK-only). */
     pricing: priceSnapshotSchema.optional(),
+    /**
+     * Kaana's price of one minute of this model's work, whatever units its
+     * provider bills (contract set 6.0.0): for a realtime model, a minute of a
+     * conversation under `server_vad`, Kaana's default turn detection; for a
+     * transcription model, a minute of audio. An estimate to reserve or show
+     * with, rounded up; what a request costs is what it measured, in `pricing`.
+     */
+    pricePerMinute: moneySchema.optional(),
     deprecation: modelDeprecationSchema,
     evaluations: z.array(modelEvaluationResultSchema).default([]),
     safety: modelSafetyMetadataSchema.optional(),

package/dist/public/messages.d.ts

@@ -250,6 +250,17 @@ export declare const inferenceSpeechParametersSchema: z.ZodObject<{
     }>;
     speed: z.ZodOptional<z.ZodNumber>;
 }, z.core.$strict>;
+/**
+ * Parameters for a file transcription, preserved in the signed request
+ * (contract set 4.5.0). `language` is the caller's statement of what is spoken,
+ * in the tag the model's catalogue `audio.languages` declares; absent lets the
+ * model detect it. An adapter whose provider takes no language refuses one
+ * rather than dropping it.
+ */
+export declare const inferenceTranscriptionParametersSchema: z.ZodObject<{
+    language: z.ZodOptional<z.ZodString>;
+    prompt: z.ZodOptional<z.ZodString>;
+}, z.core.$strict>;
 /**
  * Spoken output from a CONVERSATIONAL model (contract set 3.2.0) — an audio chat
  * model answering in its own voice, as opposed to `speech`, which reads given
@@ -283,6 +294,7 @@ export declare const inferenceAudioOutputParametersSchema: z.ZodObject<{
  */
 export declare const requestLabelsSchema: z.ZodRecord<z.ZodString, z.ZodString>;
 export type InferenceSpeechParameters = z.infer<typeof inferenceSpeechParametersSchema>;
+export type InferenceTranscriptionParameters = z.infer<typeof inferenceTranscriptionParametersSchema>;
 export type InferenceAudioOutputParameters = z.infer<typeof inferenceAudioOutputParametersSchema>;
 export type InferenceContentSource = z.infer<typeof inferenceContentSourceSchema>;
 export type InferenceContentPart = z.infer<typeof inferenceContentPartSchema>;

package/dist/public/messages.js

@@ -5,6 +5,7 @@
  * Moved from `@oxy.so/contracts` 4.12.0 (`src/inference/request.ts`) by Kaana ADR 0002.
  */
 import { z } from 'zod';
+import { languageTagSchema } from './audioFormats.js';
 import { reasoningEffortSchema } from './catalogue.js';
 /**
  * Where binary or remote content comes from.
@@ -219,6 +220,20 @@ export const inferenceSpeechParametersSchema = z
     speed: z.number().min(0.25).max(4).optional(),
 })
     .strict();
+/**
+ * Parameters for a file transcription, preserved in the signed request
+ * (contract set 4.5.0). `language` is the caller's statement of what is spoken,
+ * in the tag the model's catalogue `audio.languages` declares; absent lets the
+ * model detect it. An adapter whose provider takes no language refuses one
+ * rather than dropping it.
+ */
+export const inferenceTranscriptionParametersSchema = z
+    .object({
+    language: languageTagSchema.optional(),
+    /** Vocabulary or context, never an instruction (contract set 5.0.0). */
+    prompt: z.string().max(4096).optional(),
+})
+    .strict();
 /**
  * Spoken output from a CONVERSATIONAL model (contract set 3.2.0) — an audio chat
  * model answering in its own voice, as opposed to `speech`, which reads given

package/dist/public/realtime.d.ts

@@ -4,6 +4,7 @@
  * Moved from `@oxy.so/contracts` 4.12.0 (`src/inference/realtime.ts`) by Kaana ADR 0002.
  */
 import { z } from 'zod';
+import { realtimeOutputModalitySchema } from './audioFormats.js';
 import { type ModelAudioCapabilities, type ModelCapabilities, realtimeSessionKindSchema } from './catalogue.js';
 /**
  * The most base64 characters one audio frame may carry: 49 152 bytes of audio,
@@ -28,11 +29,6 @@ export declare const realtimeCommandIdSchema: z.ZodString;
 export declare const realtimeItemIdSchema: z.ZodString;
 /** A response's identity within one session. */
 export declare const realtimeResponseIdSchema: z.ZodString;
-/** What a session's responses produce. */
-export declare const realtimeOutputModalitySchema: z.ZodEnum<{
-    text: "text";
-    audio: "audio";
-}>;
 /**
  * How the end of a user's turn is detected.
  *
@@ -41,6 +37,10 @@ export declare const realtimeOutputModalitySchema: z.ZodEnum<{
  * 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.
+ *
+ * A `server_vad` that names no `silenceDurationMs` waits Kaana's
+ * {@link REALTIME_DEFAULT_SILENCE_DURATION_MS}, on every provider (contract set
+ * 6.0.0), never the provider's own default.
  */
 export declare const realtimeTurnDetectionSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
     type: z.ZodLiteral<"none">;
@@ -69,13 +69,12 @@ export declare const realtimeTurnDetectionSchema: z.ZodDiscriminatedUnion<[z.Zod
  *
  * 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`.
+ * and each transcribed turn arrives as deltas that only append and EXACTLY
+ * ONE `transcript.done` with `source: 'input_audio'`, the final transcript of
+ * that item (contract set 6.0.0), whatever its provider sends. 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.
  */
 export declare const realtimeInputTranscriptionSchema: z.ZodObject<{
     language: z.ZodOptional<z.ZodString>;
@@ -85,6 +84,16 @@ export declare const realtimeInputTranscriptionSchema: z.ZodObject<{
 export declare const realtimeTranslationSchema: z.ZodObject<{
     targetLanguage: z.ZodString;
 }, z.core.$strict>;
+/**
+ * A function tool a session may call: the chat tool without `strict`, which no
+ * realtime provider's session schema has (contract set 5.0.0).
+ */
+export declare const realtimeToolDefinitionSchema: z.ZodObject<{
+    type: z.ZodLiteral<"function">;
+    name: z.ZodString;
+    description: z.ZodOptional<z.ZodString>;
+    parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
+}, z.core.$strict>;
 /**
  * A session's full configuration, as signed at open and as echoed back by
  * `session.created` and `session.updated`.
@@ -143,7 +152,6 @@ export declare const realtimeSessionConfigSchema: z.ZodObject<{
         name: z.ZodString;
         description: z.ZodOptional<z.ZodString>;
         parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
-        strict: z.ZodOptional<z.ZodBoolean>;
     }, z.core.$strict>>>;
     toolChoice: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
         auto: "auto";
@@ -153,7 +161,6 @@ export declare const realtimeSessionConfigSchema: z.ZodObject<{
         type: z.ZodLiteral<"function">;
         name: z.ZodString;
     }, z.core.$strict>]>>;
-    temperature: z.ZodOptional<z.ZodNumber>;
     maxOutputTokens: z.ZodOptional<z.ZodNumber>;
 }, z.core.$strict>;
 /**
@@ -199,7 +206,6 @@ export declare const realtimeSessionConfigUpdateSchema: z.ZodObject<{
         name: z.ZodString;
         description: z.ZodOptional<z.ZodString>;
         parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
-        strict: z.ZodOptional<z.ZodBoolean>;
     }, z.core.$strict>>>;
     toolChoice: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
         auto: "auto";
@@ -209,7 +215,6 @@ export declare const realtimeSessionConfigUpdateSchema: z.ZodObject<{
         type: z.ZodLiteral<"function">;
         name: z.ZodString;
     }, z.core.$strict>]>>;
-    temperature: z.ZodOptional<z.ZodNumber>;
     maxOutputTokens: z.ZodOptional<z.ZodNumber>;
 }, z.core.$strict>;
 /**
@@ -360,7 +365,6 @@ export declare const realtimeSessionUpdateCommandSchema: z.ZodObject<{
             name: z.ZodString;
             description: z.ZodOptional<z.ZodString>;
             parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
-            strict: z.ZodOptional<z.ZodBoolean>;
         }, z.core.$strict>>>;
         toolChoice: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
             auto: "auto";
@@ -370,7 +374,6 @@ export declare const realtimeSessionUpdateCommandSchema: z.ZodObject<{
             type: z.ZodLiteral<"function">;
             name: z.ZodString;
         }, z.core.$strict>]>>;
-        temperature: z.ZodOptional<z.ZodNumber>;
         maxOutputTokens: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>;
     schemaVersion: z.ZodLiteral<1>;
@@ -552,7 +555,6 @@ export declare const realtimeClientCommandSchema: z.ZodDiscriminatedUnion<[z.Zod
             name: z.ZodString;
             description: z.ZodOptional<z.ZodString>;
             parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
-            strict: z.ZodOptional<z.ZodBoolean>;
         }, z.core.$strict>>>;
         toolChoice: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
             auto: "auto";
@@ -562,7 +564,6 @@ export declare const realtimeClientCommandSchema: z.ZodDiscriminatedUnion<[z.Zod
             type: z.ZodLiteral<"function">;
             name: z.ZodString;
         }, z.core.$strict>]>>;
-        temperature: z.ZodOptional<z.ZodNumber>;
         maxOutputTokens: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>;
     schemaVersion: z.ZodLiteral<1>;
@@ -738,7 +739,6 @@ export declare const realtimeSessionUpdatedEventSchema: z.ZodObject<{
             name: z.ZodString;
             description: z.ZodOptional<z.ZodString>;
             parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
-            strict: z.ZodOptional<z.ZodBoolean>;
         }, z.core.$strict>>>;
         toolChoice: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
             auto: "auto";
@@ -748,7 +748,6 @@ export declare const realtimeSessionUpdatedEventSchema: z.ZodObject<{
             type: z.ZodLiteral<"function">;
             name: z.ZodString;
         }, z.core.$strict>]>>;
-        temperature: z.ZodOptional<z.ZodNumber>;
         maxOutputTokens: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>;
     schemaVersion: z.ZodLiteral<1>;
@@ -1065,15 +1064,9 @@ export interface RealtimeConfigIssue {
  * apply the same ones, and a session a client may open is one Kaana may run.
  */
 export declare function realtimeConfigIssues(kind: z.infer<typeof realtimeSessionKindSchema>, config: z.infer<typeof realtimeSessionConfigSchema>): RealtimeConfigIssue[];
-/**
- * The most text `conversation.item.create` commands a session may be opened
- * for: the bound a duration-priced route's hold is sized from.
- */
-export declare const MAX_REALTIME_TEXT_ITEMS = 10000;
 /**
  * The customer's partial limits request: every signed limit, each filled from
- * Kaana's defaults when omitted, plus `maxTextItems`, the most billable text
- * items the session may send (absent, it defaults to `maxResponses`).
+ * Kaana's defaults when omitted, never above them.
  */
 export declare const realtimeOpenLimitsSchema: z.ZodObject<{
     maxDurationMs: z.ZodOptional<z.ZodNumber>;
@@ -1081,8 +1074,23 @@ export declare const realtimeOpenLimitsSchema: z.ZodObject<{
     maxInputAudioBytes: z.ZodOptional<z.ZodNumber>;
     maxOutputAudioBytes: z.ZodOptional<z.ZodNumber>;
     maxResponses: z.ZodOptional<z.ZodNumber>;
-    maxTextItems: z.ZodOptional<z.ZodNumber>;
 }, z.core.$strict>;
+/**
+ * The largest first frame of `GET /v1/realtime`: a `session.open` with its
+ * instructions and up to 128 tool definitions (contract set 5.0.0). A larger
+ * frame closes the connection before it is read.
+ */
+export declare const MAX_REALTIME_OPEN_FRAME_BYTES: number;
+/**
+ * The silence, in milliseconds, after which `server_vad` ends a caller's turn
+ * when the session names none (contract set 6.0.0, Kaana ADR 0003). Kaana
+ * sends it to every provider, so a turn ends the same way whichever serves
+ * the session. Measured on x-ai/grok-voice-3.0 (2026-10-16): with xAI's own
+ * default a phrase with pauses of 0.56 s and 0.79 s was cut into three
+ * responses, two cancelled, the model speaking before the phrase ended; at
+ * 500 ms it was one response, answered 490–576 ms after the phrase.
+ */
+export declare const REALTIME_DEFAULT_SILENCE_DURATION_MS = 500;
 /**
  * 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
@@ -1149,7 +1157,6 @@ export declare const realtimeSessionOpenSchema: z.ZodObject<{
             name: z.ZodString;
             description: z.ZodOptional<z.ZodString>;
             parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
-            strict: z.ZodOptional<z.ZodBoolean>;
         }, z.core.$strict>>>;
         toolChoice: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
             auto: "auto";
@@ -1159,7 +1166,6 @@ export declare const realtimeSessionOpenSchema: z.ZodObject<{
             type: z.ZodLiteral<"function">;
             name: z.ZodString;
         }, z.core.$strict>]>>;
-        temperature: z.ZodOptional<z.ZodNumber>;
         maxOutputTokens: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>;
     limits: z.ZodOptional<z.ZodObject<{
@@ -1168,7 +1174,6 @@ export declare const realtimeSessionOpenSchema: z.ZodObject<{
         maxInputAudioBytes: z.ZodOptional<z.ZodNumber>;
         maxOutputAudioBytes: z.ZodOptional<z.ZodNumber>;
         maxResponses: z.ZodOptional<z.ZodNumber>;
-        maxTextItems: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>>;
     clientSessionId: z.ZodOptional<z.ZodString>;
     labels: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;
@@ -1262,7 +1267,6 @@ export declare const realtimePublicSessionCreatedEventSchema: z.ZodObject<{
             name: z.ZodString;
             description: z.ZodOptional<z.ZodString>;
             parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
-            strict: z.ZodOptional<z.ZodBoolean>;
         }, z.core.$strict>>>;
         toolChoice: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
             auto: "auto";
@@ -1272,7 +1276,6 @@ export declare const realtimePublicSessionCreatedEventSchema: z.ZodObject<{
             type: z.ZodLiteral<"function">;
             name: z.ZodString;
         }, z.core.$strict>]>>;
-        temperature: z.ZodOptional<z.ZodNumber>;
         maxOutputTokens: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>;
     limits: z.ZodObject<{
@@ -1501,7 +1504,6 @@ export declare const realtimePublicServerEventSchema: z.ZodDiscriminatedUnion<[z
             name: z.ZodString;
             description: z.ZodOptional<z.ZodString>;
             parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
-            strict: z.ZodOptional<z.ZodBoolean>;
         }, z.core.$strict>>>;
         toolChoice: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
             auto: "auto";
@@ -1511,7 +1513,6 @@ export declare const realtimePublicServerEventSchema: z.ZodDiscriminatedUnion<[z
             type: z.ZodLiteral<"function">;
             name: z.ZodString;
         }, z.core.$strict>]>>;
-        temperature: z.ZodOptional<z.ZodNumber>;
         maxOutputTokens: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>;
     limits: z.ZodObject<{
@@ -1579,7 +1580,6 @@ export declare const realtimePublicServerEventSchema: z.ZodDiscriminatedUnion<[z
             name: z.ZodString;
             description: z.ZodOptional<z.ZodString>;
             parameters: z.ZodRecord<z.ZodString, z.ZodUnknown>;
-            strict: z.ZodOptional<z.ZodBoolean>;
         }, z.core.$strict>>>;
         toolChoice: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
             auto: "auto";
@@ -1589,7 +1589,6 @@ export declare const realtimePublicServerEventSchema: z.ZodDiscriminatedUnion<[z
             type: z.ZodLiteral<"function">;
             name: z.ZodString;
         }, z.core.$strict>]>>;
-        temperature: z.ZodOptional<z.ZodNumber>;
         maxOutputTokens: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strict>;
     schemaVersion: z.ZodLiteral<1>;

package/dist/public/realtime.js

@@ -4,7 +4,7 @@
  * 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 { realtimeAudioFormatSchema, realtimeOutputModalitySchema } from './audioFormats.js';
 import { realtimeSessionKindSchema, realtimeSessionTransportSchema, } from './catalogue.js';
 import { inferenceTimestampSchema, modelReferenceSchema, PADDED_BASE64_PATTERN, requestIdSchema, } from './identifiers.js';
 import { publicErrorSchema } from './errors.js';
@@ -34,8 +34,6 @@ export const realtimeCommandIdSchema = z.string().min(1).max(128);
 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)
@@ -51,6 +49,10 @@ const milliseconds = z.number().int().nonnegative().safe();
  * 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.
+ *
+ * A `server_vad` that names no `silenceDurationMs` waits Kaana's
+ * {@link REALTIME_DEFAULT_SILENCE_DURATION_MS}, on every provider (contract set
+ * 6.0.0), never the provider's own default.
  */
 export const realtimeTurnDetectionSchema = z.discriminatedUnion('type', [
     z.object({ type: z.literal('none') }).strict(),
@@ -80,18 +82,22 @@ export const realtimeTurnDetectionSchema = z.discriminatedUnion('type', [
  *
  * 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`.
+ * and each transcribed turn arrives as deltas that only append and EXACTLY
+ * ONE `transcript.done` with `source: 'input_audio'`, the final transcript of
+ * that item (contract set 6.0.0), whatever its provider sends. 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.
  */
 export const realtimeInputTranscriptionSchema = z
     .object({
-    /** A BCP 47 tag, when the caller knows the language. */
+    /**
+     * The client's own BCP 47 tag, when it knows the language. Kaana sends
+     * the provider its spelling of that language, or none for the model to
+     * detect; it is never refused.
+     */
     language: z.string().min(2).max(35).optional(),
+    /** A hint — vocabulary or context — used where the model takes one. */
     prompt: z.string().max(4_096).optional(),
 })
     .strict();
@@ -102,6 +108,11 @@ export const realtimeTranslationSchema = z
     targetLanguage: z.string().min(2).max(35),
 })
     .strict();
+/**
+ * A function tool a session may call: the chat tool without `strict`, which no
+ * realtime provider's session schema has (contract set 5.0.0).
+ */
+export const realtimeToolDefinitionSchema = toolDefinitionSchema.omit({ strict: true });
 /**
  * A session's full configuration, as signed at open and as echoed back by
  * `session.created` and `session.updated`.
@@ -121,9 +132,8 @@ export const realtimeSessionConfigSchema = z
     turnDetection: realtimeTurnDetectionSchema,
     inputAudioTranscription: realtimeInputTranscriptionSchema.optional(),
     translation: realtimeTranslationSchema.optional(),
-    tools: z.array(toolDefinitionSchema).max(128).optional(),
+    tools: z.array(realtimeToolDefinitionSchema).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(),
 })
@@ -142,9 +152,8 @@ export const realtimeSessionConfigUpdateSchema = z
     outputModalities: z.array(realtimeOutputModalitySchema).min(1).max(2).optional(),
     turnDetection: realtimeTurnDetectionSchema.optional(),
     inputAudioTranscription: realtimeInputTranscriptionSchema.optional(),
-    tools: z.array(toolDefinitionSchema).max(128).optional(),
+    tools: z.array(realtimeToolDefinitionSchema).max(128).optional(),
     toolChoice: toolChoiceSchema.optional(),
-    temperature: z.number().min(0).max(2).optional(),
     maxOutputTokens: z.number().int().positive().safe().optional(),
 })
     .strict()
@@ -587,21 +596,26 @@ export function realtimeConfigIssues(kind, config) {
 /*  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.
+ * The customer's partial limits request: every signed limit, each filled from
+ * Kaana's defaults when omitted, never above them.
  */
-export const MAX_REALTIME_TEXT_ITEMS = 10_000;
+export const realtimeOpenLimitsSchema = realtimeSessionLimitsSchema.partial().strict();
 /**
- * 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`).
+ * The largest first frame of `GET /v1/realtime`: a `session.open` with its
+ * instructions and up to 128 tool definitions (contract set 5.0.0). A larger
+ * frame closes the connection before it is read.
  */
-export const realtimeOpenLimitsSchema = realtimeSessionLimitsSchema
-    .partial()
-    .extend({
-    maxTextItems: z.number().int().nonnegative().max(MAX_REALTIME_TEXT_ITEMS).optional(),
-})
-    .strict();
+export const MAX_REALTIME_OPEN_FRAME_BYTES = 1024 * 1024;
+/**
+ * The silence, in milliseconds, after which `server_vad` ends a caller's turn
+ * when the session names none (contract set 6.0.0, Kaana ADR 0003). Kaana
+ * sends it to every provider, so a turn ends the same way whichever serves
+ * the session. Measured on x-ai/grok-voice-3.0 (2026-10-16): with xAI's own
+ * default a phrase with pauses of 0.56 s and 0.79 s was cut into three
+ * responses, two cancelled, the model speaking before the phrase ended; at
+ * 500 ms it was one response, answered 490–576 ms after the phrase.
+ */
+export const REALTIME_DEFAULT_SILENCE_DURATION_MS = 500;
 /**
  * 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
@@ -697,14 +711,6 @@ export function checkRealtimeSessionOpen(open, model) {
             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;
 }
 /* -------------------------------------------------------------------------- */

package/dist/public/version.d.ts

@@ -100,7 +100,7 @@
  * change to, say, the catalogue reject every in-flight inference request; the
  * per-shape `schemaVersion` is what a message is validated against.
  */
-export declare const INFERENCE_CONTRACT_VERSION = "4.4.0";
+export declare const INFERENCE_CONTRACT_VERSION = "6.0.0";
 /**
  * Contract set 4.0.0 (Kaana ADR 0002) is the first published as
  * `@kaana.ai/contracts`. MAJOR because the set changed hands and split in two:
@@ -135,12 +135,52 @@ export declare const INFERENCE_CONTRACT_VERSION = "4.4.0";
  * permission, dedicated capacity), and `sameModelFailover` on
  * `grantStepLimitsSchema` (default `true`: what 4.3.0 did). Nothing in 4.3.0
  * changes.
+ *
+ * Contract set 4.5.0 is additive: the signed envelope's optional
+ * `transcription` member (`inferenceTranscriptionParametersSchema`), which
+ * carries a file transcription's `language` to the model instead of the
+ * public API refusing it. Nothing in 4.4.0 changes.
+ *
+ * Contract set 5.0.0 publishes what each voice model ACCEPTS, so a client
+ * builds its request from the catalogue instead of learning the rules from a
+ * refusal. Additive: `capabilities.realtime.parameters` (turn detections and
+ * their bounds, output modality sets, tool choices, the output token cap, the
+ * input transcription prompt, and the units each detection is billed in),
+ * `audio.detectsLanguage` and `audio.transcriptionPrompt`,
+ * `realtimeParameterIssues` and `checkTranscriptionRequest`, the envelope's
+ * `transcription.prompt`, `MAX_REALTIME_OPEN_FRAME_BYTES`. MAJOR because it
+ * REMOVES what no server ever served, each of which Kaana already refused:
+ * the realtime `temperature`, a realtime tool's `strict`
+ * (`realtimeToolDefinitionSchema`), `limits.maxTextItems` and
+ * `MAX_REALTIME_TEXT_ITEMS`, and a transcription's `timestamps`, `segments`
+ * and `words`. A request that parsed under 4.5.0 and that Kaana served
+ * still parses and is served the same.
+ *
+ * Contract set 6.0.0 (Kaana ADR 0003) is one canonical API: no provider's
+ * variant reaches a client. Kaana translates or emulates every value the
+ * contract offers on every model, so the catalogue publishes product
+ * capabilities only. MAJOR because it REMOVES what 5.0.0 published of each
+ * provider's wire: `capabilities.realtime.parameters`
+ * (`realtimeParametersSchema`, `realtimeTurnDetectionParametersSchema`,
+ * `numberRangeSchema`, `realtimeToolChoiceKindSchema`),
+ * `realtimeParameterIssues`, `audio.detectsLanguage` and
+ * `audio.transcriptionPrompt`; and `checkTranscriptionRequest` takes the
+ * upload's media type alone, since a language or a prompt is never refused.
+ * Additive: a model's `pricePerMinute`, exactly one input `transcript.done`
+ * per item, and `REALTIME_DEFAULT_SILENCE_DURATION_MS`, the silence a
+ * `server_vad` that names none waits on every provider. A client on 5.0.0
+ * parses every 6.0.0 response.
  */
 /** The public API's one origin and version. */
 export declare const KAANA_API_BASE_URL = "https://kaana.ai/v1";
 /** The response header that carries the request id. */
 export declare const KAANA_REQUEST_ID_HEADER = "Kaana-Request-Id";
-/** The response header that carries {@link INFERENCE_CONTRACT_VERSION}. */
+/**
+ * The header that carries {@link INFERENCE_CONTRACT_VERSION}: on every
+ * response, the set Kaana serves; on a request (`@kaana.ai/sdk` 2.0.0 and
+ * later), the set the client parses. Every set since 4.0.0 parses the
+ * catalogue Kaana serves.
+ */
 export declare const KAANA_CONTRACT_VERSION_HEADER = "Kaana-Contract-Version";
 /**
  * The response header that names the revision that actually served a request

package/dist/public/version.js

@@ -100,7 +100,7 @@
  * change to, say, the catalogue reject every in-flight inference request; the
  * per-shape `schemaVersion` is what a message is validated against.
  */
-export const INFERENCE_CONTRACT_VERSION = '4.4.0';
+export const INFERENCE_CONTRACT_VERSION = '6.0.0';
 /**
  * Contract set 4.0.0 (Kaana ADR 0002) is the first published as
  * `@kaana.ai/contracts`. MAJOR because the set changed hands and split in two:
@@ -135,12 +135,52 @@ export const INFERENCE_CONTRACT_VERSION = '4.4.0';
  * permission, dedicated capacity), and `sameModelFailover` on
  * `grantStepLimitsSchema` (default `true`: what 4.3.0 did). Nothing in 4.3.0
  * changes.
+ *
+ * Contract set 4.5.0 is additive: the signed envelope's optional
+ * `transcription` member (`inferenceTranscriptionParametersSchema`), which
+ * carries a file transcription's `language` to the model instead of the
+ * public API refusing it. Nothing in 4.4.0 changes.
+ *
+ * Contract set 5.0.0 publishes what each voice model ACCEPTS, so a client
+ * builds its request from the catalogue instead of learning the rules from a
+ * refusal. Additive: `capabilities.realtime.parameters` (turn detections and
+ * their bounds, output modality sets, tool choices, the output token cap, the
+ * input transcription prompt, and the units each detection is billed in),
+ * `audio.detectsLanguage` and `audio.transcriptionPrompt`,
+ * `realtimeParameterIssues` and `checkTranscriptionRequest`, the envelope's
+ * `transcription.prompt`, `MAX_REALTIME_OPEN_FRAME_BYTES`. MAJOR because it
+ * REMOVES what no server ever served, each of which Kaana already refused:
+ * the realtime `temperature`, a realtime tool's `strict`
+ * (`realtimeToolDefinitionSchema`), `limits.maxTextItems` and
+ * `MAX_REALTIME_TEXT_ITEMS`, and a transcription's `timestamps`, `segments`
+ * and `words`. A request that parsed under 4.5.0 and that Kaana served
+ * still parses and is served the same.
+ *
+ * Contract set 6.0.0 (Kaana ADR 0003) is one canonical API: no provider's
+ * variant reaches a client. Kaana translates or emulates every value the
+ * contract offers on every model, so the catalogue publishes product
+ * capabilities only. MAJOR because it REMOVES what 5.0.0 published of each
+ * provider's wire: `capabilities.realtime.parameters`
+ * (`realtimeParametersSchema`, `realtimeTurnDetectionParametersSchema`,
+ * `numberRangeSchema`, `realtimeToolChoiceKindSchema`),
+ * `realtimeParameterIssues`, `audio.detectsLanguage` and
+ * `audio.transcriptionPrompt`; and `checkTranscriptionRequest` takes the
+ * upload's media type alone, since a language or a prompt is never refused.
+ * Additive: a model's `pricePerMinute`, exactly one input `transcript.done`
+ * per item, and `REALTIME_DEFAULT_SILENCE_DURATION_MS`, the silence a
+ * `server_vad` that names none waits on every provider. A client on 5.0.0
+ * parses every 6.0.0 response.
  */
 /** The public API's one origin and version. */
 export const KAANA_API_BASE_URL = 'https://kaana.ai/v1';
 /** The response header that carries the request id. */
 export const KAANA_REQUEST_ID_HEADER = 'Kaana-Request-Id';
-/** The response header that carries {@link INFERENCE_CONTRACT_VERSION}. */
+/**
+ * The header that carries {@link INFERENCE_CONTRACT_VERSION}: on every
+ * response, the set Kaana serves; on a request (`@kaana.ai/sdk` 2.0.0 and
+ * later), the set the client parses. Every set since 4.0.0 parses the
+ * catalogue Kaana serves.
+ */
 export const KAANA_CONTRACT_VERSION_HEADER = 'Kaana-Contract-Version';
 /**
  * The response header that names the revision that actually served a request

package/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "Kaana API",
-    "version": "4.4.0",
+    "version": "6.0.0",
     "description": "Kaana inference API. Generated from @kaana.ai/contracts; the Zod source is the authority. Accounts, applications, oxy_sk_ keys, limits and billing are managed in Oxy Console: https://console.oxy.so"
   },
   "servers": [
@@ -3191,6 +3191,19 @@
         ],
         "additionalProperties": false
       },
+      "InferenceTranscriptionParameters": {
+        "type": "object",
+        "properties": {
+          "language": {
+            "$ref": "#/components/schemas/LanguageTag"
+          },
+          "prompt": {
+            "type": "string",
+            "maxLength": 4096
+          }
+        },
+        "additionalProperties": false
+      },
       "LanguageTag": {
         "type": "string",
         "maxLength": 35,
@@ -3807,6 +3820,9 @@
           "pricing": {
             "$ref": "#/components/schemas/PriceSnapshot"
           },
+          "pricePerMinute": {
+            "$ref": "#/components/schemas/Money"
+          },
           "deprecation": {
             "$ref": "#/components/schemas/ModelDeprecation"
           },
@@ -4760,11 +4776,6 @@
             "type": "integer",
             "exclusiveMinimum": 0,
             "maximum": 10000
-          },
-          "maxTextItems": {
-            "type": "integer",
-            "minimum": 0,
-            "maximum": 10000
           }
         },
         "additionalProperties": false
@@ -5352,17 +5363,12 @@
             "maxItems": 128,
             "type": "array",
             "items": {
-              "$ref": "#/components/schemas/ToolDefinition"
+              "$ref": "#/components/schemas/RealtimeToolDefinition"
             }
           },
           "toolChoice": {
             "$ref": "#/components/schemas/ToolChoice"
           },
-          "temperature": {
-            "type": "number",
-            "minimum": 0,
-            "maximum": 2
-          },
           "maxOutputTokens": {
             "type": "integer",
             "exclusiveMinimum": 0,
@@ -5400,17 +5406,12 @@
             "maxItems": 128,
             "type": "array",
             "items": {
-              "$ref": "#/components/schemas/ToolDefinition"
+              "$ref": "#/components/schemas/RealtimeToolDefinition"
             }
           },
           "toolChoice": {
             "$ref": "#/components/schemas/ToolChoice"
           },
-          "temperature": {
-            "type": "number",
-            "minimum": 0,
-            "maximum": 2
-          },
           "maxOutputTokens": {
             "type": "integer",
             "exclusiveMinimum": 0,
@@ -5820,6 +5821,37 @@
           "complete"
         ]
       },
+      "RealtimeToolDefinition": {
+        "type": "object",
+        "properties": {
+          "type": {
+            "type": "string",
+            "const": "function"
+          },
+          "name": {
+            "type": "string",
+            "minLength": 1,
+            "maxLength": 128
+          },
+          "description": {
+            "type": "string",
+            "maxLength": 2000
+          },
+          "parameters": {
+            "type": "object",
+            "propertyNames": {
+              "type": "string"
+            },
+            "additionalProperties": {}
+          }
+        },
+        "required": [
+          "type",
+          "name",
+          "parameters"
+        ],
+        "additionalProperties": false
+      },
       "RealtimeTranscriptDeltaEvent": {
         "type": "object",
         "properties": {
@@ -6682,14 +6714,6 @@
             "type": "string",
             "maxLength": 4096
           },
-          "timestamps": {
-            "type": "string",
-            "enum": [
-              "none",
-              "segment",
-              "word"
-            ]
-          },
           "audio": {
             "type": "object",
             "properties": {
@@ -6728,14 +6752,6 @@
           "prompt": {
             "type": "string",
             "maxLength": 4096
-          },
-          "timestamps": {
-            "type": "string",
-            "enum": [
-              "none",
-              "segment",
-              "word"
-            ]
           }
         },
         "required": [
@@ -6767,60 +6783,6 @@
             "minimum": 0,
             "maximum": 9007199254740991
           },
-          "segments": {
-            "type": "array",
-            "items": {
-              "type": "object",
-              "properties": {
-                "startMs": {
-                  "type": "integer",
-                  "minimum": 0,
-                  "maximum": 9007199254740991
-                },
-                "endMs": {
-                  "type": "integer",
-                  "minimum": 0,
-                  "maximum": 9007199254740991
-                },
-                "text": {
-                  "type": "string"
-                }
-              },
-              "required": [
-                "startMs",
-                "endMs",
-                "text"
-              ],
-              "additionalProperties": false
-            }
-          },
-          "words": {
-            "type": "array",
-            "items": {
-              "type": "object",
-              "properties": {
-                "startMs": {
-                  "type": "integer",
-                  "minimum": 0,
-                  "maximum": 9007199254740991
-                },
-                "endMs": {
-                  "type": "integer",
-                  "minimum": 0,
-                  "maximum": 9007199254740991
-                },
-                "word": {
-                  "type": "string"
-                }
-              },
-              "required": [
-                "startMs",
-                "endMs",
-                "word"
-              ],
-              "additionalProperties": false
-            }
-          },
           "usage": {
             "$ref": "#/components/schemas/Usage"
           }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kaana.ai/contracts",
-  "version": "4.4.1",
+  "version": "6.0.0",
   "description": "Kaana inference contract: the Zod schemas of the public API at https://kaana.ai/v1 and, under /internal, of the Oxy-Kaana grant, settlement and signed lane.",
   "type": "module",
   "sideEffects": false,