@oxy.so/contracts 4.3.0 → 4.4.0

package/dist/cjs/inference/request.js

@@ -25,7 +25,7 @@
  * docs/adr/0017-authorized-routes-in-the-envelope.md.
  */
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.inferenceRequestSchema = exports.inferenceSpeechParametersSchema = exports.clientRequestMetadataSchema = exports.inferenceReasoningSchema = exports.responseFormatSchema = exports.toolChoiceSchema = exports.toolDefinitionSchema = exports.samplingParametersSchema = exports.inferenceInputSchema = exports.inferenceMessageSchema = exports.inferenceMessageRoleSchema = exports.inferenceToolCallSchema = exports.inferenceContentPartSchema = exports.inferenceContentSourceSchema = void 0;
+exports.inferenceRequestSchema = exports.inferenceAudioOutputParametersSchema = exports.inferenceSpeechParametersSchema = exports.clientRequestMetadataSchema = exports.inferenceReasoningSchema = exports.responseFormatSchema = exports.toolChoiceSchema = exports.toolDefinitionSchema = exports.samplingParametersSchema = exports.inferenceInputSchema = exports.inferenceMessageSchema = exports.inferenceMessageRoleSchema = exports.inferenceToolCallSchema = exports.inferenceContentPartSchema = exports.inferenceContentSourceSchema = void 0;
 const zod_1 = require("zod");
 const attribution_1 = require("./attribution");
 const catalogue_1 = require("./catalogue");
@@ -292,16 +292,7 @@ exports.inferenceReasoningSchema = zod_1.z
 exports.clientRequestMetadataSchema = zod_1.z
     .object({
     /** The public dialect the customer called. The response is rendered in it. */
-    apiFormat: zod_1.z.enum([
-        "responses",
-        "chat_completions",
-        "embeddings",
-        "images_generations",
-        "audio_transcriptions",
-        "audio_speech",
-        "rerank",
-        "batches",
-    ]),
+    apiFormat: catalogue_1.inferenceApiFormatSchema,
     /** The public path, e.g. `/v1/responses`. */
     endpoint: zod_1.z.string().min(1).max(256),
     /** The customer's own correlation id, when they sent one. */
@@ -330,6 +321,25 @@ exports.inferenceSpeechParametersSchema = zod_1.z.object({
     responseFormat: zod_1.z.enum(["mp3", "opus", "aac", "flac", "wav", "pcm"]),
     speed: zod_1.z.number().min(0.25).max(4).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
+ * text aloud. Presence is the request: a chat request carrying this asks for
+ * the answer as audio AND its transcript, which is the only combination the
+ * audio chat models produce (OpenAI's `modalities: ["text", "audio"]`).
+ *
+ * The audio streams as `audio` events and the words as `delta` events on the
+ * `output_audio_transcript` channel — never as `output_text`, because a
+ * transcript of speech the customer is already hearing is not a second answer.
+ * `pcm` is the only format an audio chat model can STREAM; the refinement on the
+ * envelope holds that rather than leaving a provider to refuse it mid-flight.
+ */
+exports.inferenceAudioOutputParametersSchema = zod_1.z
+    .object({
+    voice: zod_1.z.string().min(1).max(64),
+    format: zod_1.z.enum(["wav", "mp3", "flac", "opus", "pcm"]),
+})
+    .strict();
 /**
  * The canonical internal request Oxy forwards to the data plane.
  *
@@ -352,6 +362,8 @@ exports.inferenceRequestSchema = zod_1.z
     /** Absent means the route's own default reasoning behaviour. */
     reasoning: exports.inferenceReasoningSchema.optional(),
     speech: exports.inferenceSpeechParametersSchema.optional(),
+    /** Spoken output from a conversational model. Contract set 3.2.0. */
+    audioOutput: exports.inferenceAudioOutputParametersSchema.optional(),
     tools: zod_1.z.array(exports.toolDefinitionSchema).default([]),
     toolChoice: exports.toolChoiceSchema.optional(),
     responseFormat: exports.responseFormatSchema.optional(),
@@ -401,6 +413,17 @@ exports.inferenceRequestSchema = zod_1.z
     if (!isSpeech && request.speech !== undefined) {
         ctx.addIssue({ code: zod_1.z.ZodIssueCode.custom, path: ["speech"], message: "speech parameters require the audio_speech API format" });
     }
+    if (request.audioOutput !== undefined) {
+        if (request.client.apiFormat !== "chat_completions") {
+            ctx.addIssue({ code: zod_1.z.ZodIssueCode.custom, path: ["audioOutput"], message: "spoken output from a conversational model requires the chat_completions API format" });
+        }
+        if (request.modality !== "audio" || request.input.format !== "messages") {
+            ctx.addIssue({ code: zod_1.z.ZodIssueCode.custom, path: ["audioOutput"], message: "spoken output requires audio modality and a messages input" });
+        }
+        if (request.stream && request.audioOutput.format !== "pcm") {
+            ctx.addIssue({ code: zod_1.z.ZodIssueCode.custom, path: ["audioOutput", "format"], message: "streamed spoken output is pcm; the other formats exist only whole" });
+        }
+    }
     if (request.toolChoice !== undefined && request.tools.length === 0) {
         ctx.addIssue({
             code: zod_1.z.ZodIssueCode.custom,

package/dist/cjs/inference/streamEvents.js

@@ -62,7 +62,13 @@ exports.inferenceStreamDeltaEventSchema = zod_1.z.object({
     sequence: zod_1.z.number().int().nonnegative().safe(),
     /** Which output of a multi-output response this chunk belongs to. */
     outputIndex: zod_1.z.number().int().nonnegative().safe(),
-    channel: zod_1.z.enum(['output_text', 'reasoning', 'refusal']),
+    /**
+     * `output_audio_transcript` (contract set 3.2.0) carries the words of spoken
+     * output requested with `audioOutput`. It is its own channel for the same
+     * reason `reasoning` is: a renderer that shows it as `output_text` presents
+     * the transcript of what the customer is hearing as a second, written answer.
+     */
+    channel: zod_1.z.enum(['output_text', 'reasoning', 'refusal', 'output_audio_transcript']),
     text: zod_1.z.string(),
 });
 /** The encodings an audio output may carry. */

package/dist/cjs/inference/version.js

@@ -102,4 +102,4 @@ exports.INFERENCE_CONTRACT_VERSION = void 0;
  * change to, say, the catalogue reject every in-flight inference request; the
  * per-shape `schemaVersion` is what a message is validated against.
  */
-exports.INFERENCE_CONTRACT_VERSION = '3.1.0';
+exports.INFERENCE_CONTRACT_VERSION = '3.2.0';