@juspay/neurolink 13.4.10 → 13.5.0

package/dist/cli/commands/decide.js

@@ -23,7 +23,7 @@ import { NeuroLink } from "../../neurolink.js";
 import { readDecisionChoice } from "../../utils/decisionAnswers.js";
 import { calculateCost, hasPricing } from "../../utils/pricing.js";
 import { logger } from "../../utils/logger.js";
-const VALID_QUESTION_TYPES = new Set(["boolean", "choice", "score"]);
+const VALID_QUESTION_TYPES = new Set(["boolean", "choice", "score", "value"]);
 function isRecord(value) {
     return typeof value === "object" && value !== null && !Array.isArray(value);
 }
@@ -123,6 +123,10 @@ function formatAnswerLine(id, answers) {
         }
         case "score":
             return `${id}: score ${formatNumber(answer.score)}, confidence ${formatNumber(answer.confidence)}`;
+        case "value":
+            return answer.value === null
+                ? `${id}: no value found, confidence ${formatNumber(answer.confidence)}`
+                : `${id}: value ${JSON.stringify(answer.value)}, confidence ${formatNumber(answer.confidence)}`;
         default:
             return `${id}: (unrecognised answer type)`;
     }
@@ -195,12 +199,17 @@ export const decideCommand = {
         type: "string",
         array: true,
         nargs: 1,
-        describe: "Image for the model to read: a file path or a data: URL. Repeat for several (XOR, Perplexity and OpenAI Decisions take up to 8, Cloudflare Clef up to 4)",
+        describe: "Image for the model to read: a file path or a data: URL. Repeat for several (Perplexity and OpenAI Decisions take up to 8, Cloudflare Clef up to 4, XOR 1)",
     })
         .option("video", {
         type: "string",
         nargs: 1,
-        describe: "One video for the model to read: a file path or a data: URL",
+        describe: "One video for the model to read: a file path or a data: URL. No current decision provider reads video, so it is refused before any request",
+    })
+        .option("audio", {
+        type: "string",
+        nargs: 1,
+        describe: "One audio clip for the model to hear: a WAV, FLAC, Ogg or MP3 file path or a data: URL (XOR with its xor-omni model; the first 30 seconds are heard). State may then be omitted",
     })
         .option("timeout", {
         type: "number",
@@ -221,7 +230,8 @@ export const decideCommand = {
     })
         .example('$0 decide "Refund request for a damaged item" --questions \'{"urgent":{"type":"boolean","instructions":"Is this urgent?"}}\'', "Ask a single yes/no question")
         .example("$0 decide --state-file ticket.json --questions-file questions.json --format json", "Read state and questions from files, emit raw JSON")
-        .example('$0 decide "What color is this?" --provider xor --model xor-omni --image ./photo.png --questions \'{"color":{"type":"choice","instructions":"What color is the image?","criteria":{"red":"red","blue":"blue"}}}\'', "Ask about an image (XOR with its xor-omni model, Perplexity, Cloudflare Clef and OpenAI Decisions read images)"),
+        .example('$0 decide "What color is this?" --provider xor --model xor-omni --image ./photo.png --questions \'{"color":{"type":"choice","instructions":"What color is the image?","criteria":{"red":"red","blue":"blue"}}}\'', "Ask about an image (XOR with its xor-omni model, Perplexity, Cloudflare Clef and OpenAI Decisions read images)")
+        .example('$0 decide "Invoice INV-2024-0892 for 18,275 rupees" --provider xor --model xor-omni --questions \'{"amount":{"type":"value","dtype":"int","instructions":"Invoice amount?"}}\'', "Extract a typed value (XOR)"),
     handler: async (argv) => {
         const outputFormat = argv.format ?? "text";
         // --- Validate before any provider work ---------------------------------
@@ -256,8 +266,13 @@ export const decideCommand = {
             logger.error(chalk.red("Error: pass at most one --video."));
             process.exit(1);
         }
+        const audios = [].concat(argv.audio ?? []);
+        if (audios.length > 1) {
+            logger.error(chalk.red("Error: pass at most one --audio."));
+            process.exit(1);
+        }
         if (!isDecisionQuestionMap(parsedQuestions)) {
-            logger.error(chalk.red('Error: questions must be a non-empty object of { type: "boolean" | "choice" | "score", ... }.'));
+            logger.error(chalk.red('Error: questions must be a non-empty object of { type: "boolean" | "choice" | "score" | "value", ... }.'));
             process.exit(1);
         }
         let state;
@@ -268,6 +283,10 @@ export const decideCommand = {
             logger.error(chalk.red(`Error: could not read state file — ${error instanceof Error ? error.message : String(error)}`));
             process.exit(1);
         }
+        // An audio clip can be the whole input.
+        if (state === undefined && audios.length === 1) {
+            state = "";
+        }
         if (state === undefined) {
             logger.error(chalk.red("Error: provide state as a positional argument or via --state-file."));
             process.exit(1);
@@ -284,6 +303,7 @@ export const decideCommand = {
                 timeoutMs: argv.timeout,
                 ...(argv.image && argv.image.length > 0 ? { images: argv.image } : {}),
                 ...(videos.length === 1 ? { video: videos[0] } : {}),
+                ...(audios.length === 1 ? { audio: audios[0] } : {}),
             });
             spinner?.succeed("Decision complete");
             if (outputFormat === "json") {

package/dist/factories/providerDescriptors.d.ts

@@ -45,12 +45,13 @@ export declare const DECISION_PROVIDERS: readonly ProviderDescriptor[];
  */
 export declare function resolveDefaultDecisionProvider(credentials?: NeurolinkCredentials): string | undefined;
 /**
- * Names of the decision providers that accept images, or, with `video`, the
- * ones that accept video too: a hint that names a provider which would refuse
- * the same request is worse than none.
+ * Names of the decision providers that accept images, or, with `video` or
+ * `audio`, the ones that accept that too: a hint that names a provider which
+ * would refuse the same request is worse than none.
  */
 export declare function listMediaDecisionProviders(needs?: {
     video?: boolean;
+    audio?: boolean;
 }): string[];
 /**
  * The variables that would configure each decision provider, one clause per

package/dist/factories/providerDescriptors.js

@@ -909,14 +909,16 @@ export function resolveDefaultDecisionProvider(credentials) {
     return DECISION_PROVIDERS.find((descriptor) => isDecisionProviderConfigured(descriptor, credentials))?.name;
 }
 /**
- * Names of the decision providers that accept images, or, with `video`, the
- * ones that accept video too: a hint that names a provider which would refuse
- * the same request is worse than none.
+ * Names of the decision providers that accept images, or, with `video` or
+ * `audio`, the ones that accept that too: a hint that names a provider which
+ * would refuse the same request is worse than none.
  */
 export function listMediaDecisionProviders(needs) {
     return DECISION_PROVIDERS.filter((descriptor) => {
         const media = descriptor.decisionLimits?.media;
-        return media !== undefined && (!needs?.video || media.video);
+        return (media !== undefined &&
+            (!needs?.video || media.video) &&
+            (!needs?.audio || media.audio !== undefined));
     }).map((descriptor) => descriptor.name);
 }
 /** "a", "a or b", "a, b or c". */

package/dist/index.d.ts

@@ -416,7 +416,7 @@ export { classifyProviderError, ModelPool, createDefaultRequestRouter, Classifie
  * if (team && team.confidence > 0.7) { route(team.choice); }
  * ```
  */
-export { decisionBooleanConfidence, decisionKey, gateDecisionBoolean, HOST_DECISION_NAMESPACE, isDecisionQuestion, readDecisionBoolean, readDecisionChoice, readDecisionScore, servesInferenceKind, } from "./utils/decisionAnswers.js";
+export { decisionBooleanConfidence, decisionKey, gateDecisionBoolean, HOST_DECISION_NAMESPACE, isDecisionQuestion, readDecisionBoolean, readDecisionChoice, readDecisionScore, readDecisionValue, servesInferenceKind, } from "./utils/decisionAnswers.js";
 export { estimateDecisionStateTokens, resolveDecisionLimitsReading, } from "./utils/decisionLimits.js";
 export { DECISION_PROVIDERS, resolveDefaultDecisionProvider, } from "./factories/providerDescriptors.js";
 export { TYPESAFE_MAX_REQUEST_TOKENS, TYPESAFE_MAX_STATE_TOKENS, } from "./providers/typesafe.js";

package/dist/index.js

@@ -633,7 +633,7 @@ export { classifyProviderError, ModelPool, createDefaultRequestRouter, Classifie
  * if (team && team.confidence > 0.7) { route(team.choice); }
  * ```
  */
-export { decisionBooleanConfidence, decisionKey, gateDecisionBoolean, HOST_DECISION_NAMESPACE, isDecisionQuestion, readDecisionBoolean, readDecisionChoice, readDecisionScore, servesInferenceKind, } from "./utils/decisionAnswers.js";
+export { decisionBooleanConfidence, decisionKey, gateDecisionBoolean, HOST_DECISION_NAMESPACE, isDecisionQuestion, readDecisionBoolean, readDecisionChoice, readDecisionScore, readDecisionValue, servesInferenceKind, } from "./utils/decisionAnswers.js";
 // The estimator the pre-flight refusal uses, and the flattening behind
 // `neurolink.decisionLimits()` — exported so a host measures a state with the
 // very function that will judge it.

package/dist/neurolink.d.ts

@@ -2694,12 +2694,14 @@ export declare class NeuroLink {
      *    many would have the WHOLE request refused, and NeuroLink's own routing
      *    would silently degrade to the heuristic — the host's problem would
      *    become NeuroLink's. A provider that declares no `maxQuestions`
-     *    (TypeSafe, XOR) takes every well-formed addition.
+     *    (TypeSafe) takes every well-formed addition. For the same reason a
+     *    question the provider cannot answer (a type it does not list, a choice
+     *    or score past its cap) is dropped on its own.
      * 4. Answers are split: the consumer sees only its own ids, exactly as it
      *    would without hooks; `onAnswers` gets the host's answers under the
      *    host's original ids plus the full result.
      *
-     * The request's `images` and `video` ride along untouched — `merged` and
+     * The request's `images`, `video` and `audio` ride along untouched — `merged` and
      * the wire call spread the request, and only `state`, the questions and
      * the result are cloned — so an XOR request keeps its Buffers and its
      * result keeps `mediaBytes`.

package/dist/neurolink.js

@@ -37,7 +37,7 @@ import { describeDecisionProviderKeys, PROVIDER_ALIAS_INDEX, PROVIDER_DESCRIPTOR
 import { describeSTTProviderKeys, getSTTProviderDescriptor, listConfiguredSTTProviders, resolveCallSTTEndpoints, resolveDefaultSTTProvider, resolveSTTProviderName, } from "./factories/sttDescriptors.js";
 import { MAX_PARALLEL_DECISION_BATCHES, mergeDecisionResults, splitDecisionRequest, } from "./utils/decisionBatches.js";
 import { decisionKey, HOST_DECISION_NAMESPACE, isHostDecisionKey, snapshotDecisionQuestion, stripHostAnswers, } from "./utils/decisionAnswers.js";
-import { resolveDecisionLimitsReading } from "./utils/decisionLimits.js";
+import { resolveDecisionLimitsReading, unsupportedDecisionQuestion, } from "./utils/decisionLimits.js";
 import { isDirectTTSRequest, resolveRequestKind, } from "./core/resolveRequestKind.js";
 import { createToolEventPayload } from "./core/toolEvents.js";
 import { ProviderFactory } from "./factories/providerFactory.js";
@@ -14046,12 +14046,14 @@ Current user's request: ${currentInput}`;
      *    many would have the WHOLE request refused, and NeuroLink's own routing
      *    would silently degrade to the heuristic — the host's problem would
      *    become NeuroLink's. A provider that declares no `maxQuestions`
-     *    (TypeSafe, XOR) takes every well-formed addition.
+     *    (TypeSafe) takes every well-formed addition. For the same reason a
+     *    question the provider cannot answer (a type it does not list, a choice
+     *    or score past its cap) is dropped on its own.
      * 4. Answers are split: the consumer sees only its own ids, exactly as it
      *    would without hooks; `onAnswers` gets the host's answers under the
      *    host's original ids plus the full result.
      *
-     * The request's `images` and `video` ride along untouched — `merged` and
+     * The request's `images`, `video` and `audio` ride along untouched — `merged` and
      * the wire call spread the request, and only `state`, the questions and
      * the result are cloned — so an XOR request keeps its Buffers and its
      * result keeps `mediaBytes`.
@@ -14145,7 +14147,7 @@ Current user's request: ${currentInput}`;
                         model: merged.model,
                         credentials: merged.credentials,
                     });
-                    // A provider without a question-count cap (TypeSafe, XOR) takes
+                    // A provider without a question-count cap (TypeSafe) takes
                     // every well-formed addition; its token/byte limits still apply
                     // and are what `decisionLimits()` reports for sizing.
                     const room = limits?.maxQuestions !== undefined
@@ -14155,6 +14157,7 @@ Current user's request: ${currentInput}`;
                         ...ownQuestions,
                     };
                     let invalidCount = 0;
+                    let unsupportedCount = 0;
                     let overflowCount = 0;
                     for (const [hostId, proposal] of proposedEntries) {
                         // The wire carries a snapshot of the host's question, never the
@@ -14164,6 +14167,12 @@ Current user's request: ${currentInput}`;
                             invalidCount += 1;
                             continue;
                         }
+                        // A question the provider cannot read would have the whole
+                        // request refused, NeuroLink's own questions with it.
+                        if (unsupportedDecisionQuestion(question, limits) !== undefined) {
+                            unsupportedCount += 1;
+                            continue;
+                        }
                         if (hostIdByWireId.size >= room) {
                             overflowCount += 1;
                             continue;
@@ -14175,6 +14184,9 @@ Current user's request: ${currentInput}`;
                     if (invalidCount > 0) {
                         logger.warn(`[Decision] dropped ${invalidCount} malformed host question(s) for site "${site}"`);
                     }
+                    if (unsupportedCount > 0) {
+                        logger.warn(`[Decision] dropped ${unsupportedCount} host question(s) for site "${site}" that ${limits?.provider ?? "the decision provider"} cannot answer`);
+                    }
                     if (overflowCount > 0 && limits) {
                         logger.warn(`[Decision] dropped ${overflowCount} host question(s) for site "${site}": ${limits.provider} accepts ${limits.maxQuestions} per request and NeuroLink asks ${ownCount}`);
                     }
@@ -14306,6 +14318,7 @@ Current user's request: ${currentInput}`;
                 "ai.provider": providerName,
                 "decision.question_count": questionCount,
                 "decision.images.count": options.images?.length ?? 0,
+                "decision.audio.count": options.audio === undefined ? 0 : 1,
                 ...siteSpanAttributes,
             },
         }, async (otelSpan) => {
@@ -14335,6 +14348,7 @@ Current user's request: ${currentInput}`;
                 "ai.model": options.model ?? "",
                 "decision.question_count": questionCount,
                 "decision.images.count": options.images?.length ?? 0,
+                "decision.audio.count": options.audio === undefined ? 0 : 1,
                 ...siteSpanAttributes,
             }, parentSpanId, traceId);
             try {
@@ -14348,6 +14362,7 @@ Current user's request: ${currentInput}`;
                     questions: options.questions,
                     images: options.images,
                     video: options.video,
+                    audio: options.audio,
                     model: options.model,
                     signal: options.signal,
                     timeoutMs: options.timeoutMs,

package/dist/providers/catalog/provider-catalog.schema.json

@@ -55,6 +55,25 @@
       "properties": {
         "maxStateTokens": { "type": "integer", "minimum": 1 },
         "maxQuestions": { "type": "integer", "minimum": 1 },
+        "questionTypes": {
+          "type": "array",
+          "minItems": 1,
+          "uniqueItems": true,
+          "items": {
+            "enum": ["boolean", "choice", "score", "value", "choice-unknown"]
+          },
+          "description": "The question types the provider reads; choice-unknown is a choice that sets unknown: true. Absent means boolean, choice and score. openai-decisions may list only boolean, choice and score."
+        },
+        "maxScoreLevels": {
+          "type": "integer",
+          "minimum": 1,
+          "description": "Most levels one score question may have."
+        },
+        "maxChoiceOptions": {
+          "type": "integer",
+          "minimum": 1,
+          "description": "Most options one choice question may have."
+        },
         "nonAsciiTokensPerChar": { "type": "number", "exclusiveMinimum": 0 },
         "digitTokensPerChar": { "type": "number", "exclusiveMinimum": 0 },
         "symbolTokensPerChar": { "type": "number", "exclusiveMinimum": 0 },
@@ -75,6 +94,19 @@
               "type": "integer",
               "minimum": 1,
               "description": "Ceiling on the encoded request body, in bytes."
+            },
+            "audio": {
+              "type": "object",
+              "additionalProperties": false,
+              "required": ["maxSeconds"],
+              "description": "One audio clip is accepted. Requires decide.media.audioKey; not valid with openai-decisions.",
+              "properties": {
+                "maxSeconds": {
+                  "type": "integer",
+                  "minimum": 1,
+                  "description": "Seconds of audio the model reads; a WAV, FLAC or Ogg clip that runs longer is refused."
+                }
+              }
             }
           }
         },
@@ -154,10 +186,32 @@
           "type": "object",
           "additionalProperties": false,
           "required": ["imagesKey"],
-          "description": "Body keys prepared media is written under, after model, state and questions: imagesKey only when the request carries an image, videoKey only when it carries a video, and the two differ. With dialect system-one it requires decisionLimits.media, and videoKey is required exactly when that allowance's video is true. Not valid with openai-decisions.",
+          "description": "Body keys prepared media is written under, after model, state and questions: imagesKey only when the request carries an image, videoKey only when it carries a video, audioKey only when it carries audio, and all differ. With dialect system-one it requires decisionLimits.media; videoKey is required exactly when that allowance's video is true, and audioKey exactly when it sets audio. Not valid with openai-decisions.",
           "properties": {
             "imagesKey": { "$ref": "#/definitions/decideMediaKey" },
-            "videoKey": { "$ref": "#/definitions/decideMediaKey" }
+            "videoKey": { "$ref": "#/definitions/decideMediaKey" },
+            "audioKey": { "$ref": "#/definitions/decideMediaKey" }
+          }
+        },
+        "capableModel": {
+          "type": "object",
+          "additionalProperties": false,
+          "required": ["model", "questionTypes"],
+          "description": "A models.catalog key (not the default) the request is sent to instead of the default when the call names no model, no model variable is set, and the request carries media or a question of one of questionTypes; that model's decisionLimits row is the one checked. questionTypes must be listed in decisionLimits.questionTypes.",
+          "properties": {
+            "model": { "type": "string", "minLength": 1 },
+            "questionTypes": {
+              "type": "array",
+              "items": {
+                "enum": [
+                  "boolean",
+                  "choice",
+                  "score",
+                  "value",
+                  "choice-unknown"
+                ]
+              }
+            }
           }
         },
         "ambientKeyGuard": {
@@ -570,6 +624,11 @@
           "type": "integer",
           "minimum": 1,
           "description": "Decide entries only: the per-call budget for decide(), in milliseconds. When omitted, the base provider's default timeout applies."
+        },
+        "decideMsPerKStateTokens": {
+          "type": "integer",
+          "minimum": 1,
+          "description": "Decide entries only: added to decideMs for every 1,000 estimated state tokens, for a model whose latency grows with the state."
         }
       }
     },
@@ -690,7 +749,9 @@
           "defaultMaxOutputTokens"
         ]
       },
-      "timeouts": { "properties": { "decideMs": false } }
+      "timeouts": {
+        "properties": { "decideMs": false, "decideMsPerKStateTokens": false }
+      }
     }
   },
   "definitions": {

package/dist/providers/catalog/schema.js

@@ -540,11 +540,30 @@ const decideMediaSchema = z
     .strictObject({
     imagesKey: decideMediaKeySchema,
     videoKey: decideMediaKeySchema.optional(),
+    audioKey: decideMediaKeySchema.optional(),
 })
     .refine((media) => media.imagesKey !== media.videoKey, {
     path: ["videoKey"],
     message: "decide.media.videoKey must differ from imagesKey: one body key cannot carry both",
+})
+    .refine((media) => media.audioKey === undefined ||
+    (media.audioKey !== media.imagesKey && media.audioKey !== media.videoKey), {
+    path: ["audioKey"],
+    message: "decide.media.audioKey must differ from imagesKey and videoKey: one body key cannot carry two kinds of media",
 });
+const DECIDE_QUESTION_TYPES = [
+    "boolean",
+    "choice",
+    "score",
+    "value",
+    "choice-unknown",
+];
+// The openai-decisions dialect writes predicates, choices and levels only.
+const OPENAI_DECISIONS_QUESTION_TYPES = new Set([
+    "boolean",
+    "choice",
+    "score",
+]);
 const GUARD_LIST_MAX = 8;
 const ambientGuardEnvNameSchema = z
     .string()
@@ -603,6 +622,12 @@ const decideBlockSchema = z.strictObject({
     authHeader: z.enum(["bearer", "x-api-key"]).optional(),
     stateEncoding: z.enum(["text"]).optional(),
     media: decideMediaSchema.optional(),
+    capableModel: z
+        .strictObject({
+        model: z.string().min(1),
+        questionTypes: z.array(z.enum(DECIDE_QUESTION_TYPES)),
+    })
+        .optional(),
     ambientKeyGuard: ambientKeyGuardSchema.optional(),
     requestIdHeaders: z.array(decideHeaderNameSchema).optional(),
     response: z
@@ -653,6 +678,15 @@ const decideModelsSchema = z
 const decideLimitsSchema = z.strictObject({
     maxStateTokens: decidePositiveInt,
     maxQuestions: decidePositiveInt.optional(),
+    questionTypes: z
+        .array(z.enum(DECIDE_QUESTION_TYPES))
+        .min(1)
+        .refine((types) => new Set(types).size === types.length, {
+        message: "decisionLimits.questionTypes lists the same type twice",
+    })
+        .optional(),
+    maxScoreLevels: decidePositiveInt.optional(),
+    maxChoiceOptions: decidePositiveInt.optional(),
     nonAsciiTokensPerChar: decidePositiveNumber.optional(),
     digitTokensPerChar: decidePositiveNumber.optional(),
     symbolTokensPerChar: decidePositiveNumber.optional(),
@@ -668,10 +702,36 @@ const decideLimitsSchema = z.strictObject({
         maxImages: decidePositiveInt,
         video: z.boolean(),
         maxRequestBytes: decidePositiveInt,
+        audio: z.strictObject({ maxSeconds: decidePositiveInt }).optional(),
     })
         .optional(),
     advisory: z.boolean().optional(),
 });
+// decide.capableModel names another key of the catalog, never the default, and
+// only question types the entry reads: anything else is refused before a model
+// is chosen.
+function checkCapableModel(entry, modelKeys, issue) {
+    const capable = entry.decide.capableModel;
+    if (capable === undefined) {
+        return;
+    }
+    if (!modelKeys.has(capable.model)) {
+        issue(["decide", "capableModel", "model"], "decide.capableModel.model must be a key of models.catalog");
+    }
+    if (capable.model === entry.models.default) {
+        issue(["decide", "capableModel", "model"], "decide.capableModel.model must differ from models.default: a request is sent to it instead of the default");
+    }
+    const declared = entry.decisionLimits?.questionTypes ?? [
+        "boolean",
+        "choice",
+        "score",
+    ];
+    capable.questionTypes.forEach((type, index) => {
+        if (!declared.includes(type)) {
+            issue(["decide", "capableModel", "questionTypes", index], `decide.capableModel.questionTypes lists ${type}, which decisionLimits.questionTypes does not: such a question is refused before any model is chosen`);
+        }
+    });
+}
 const decideCatalogJsonSchema = z
     .strictObject({
     ...catalogIdentityShape,
@@ -680,7 +740,10 @@ const decideCatalogJsonSchema = z
     wire: decideWireSchema,
     models: decideModelsSchema,
     timeouts: z
-        .strictObject({ decideMs: decidePositiveInt.optional() })
+        .strictObject({
+        decideMs: decidePositiveInt.optional(),
+        decideMsPerKStateTokens: decidePositiveInt.optional(),
+    })
         .optional(),
     decisionLimits: decideLimitsSchema.optional(),
     decide: decideBlockSchema,
@@ -743,6 +806,7 @@ const decideCatalogJsonSchema = z
     if (alias !== undefined && !modelKeys.has(alias)) {
         issue(["decide", "model", "alias"], "decide.model.alias must be a key of models.catalog");
     }
+    checkCapableModel(entry, modelKeys, issue);
     for (const id of Object.keys(entry.decisionLimits?.models ?? {})) {
         if (!modelKeys.has(id)) {
             issue(["decisionLimits", "models", id], "decisionLimits.models key must be a key of models.catalog");
@@ -763,6 +827,14 @@ const decideCatalogJsonSchema = z
         if (placement !== "body") {
             issue(["decide", "model", "placement"], "decide.model.placement must be body with dialect openai-decisions: the model is always a field of the request");
         }
+        if (mediaAllowance?.audio !== undefined) {
+            issue(["decisionLimits", "media", "audio"], "decisionLimits.media.audio is not valid with dialect openai-decisions: its request input has no audio part");
+        }
+        (entry.decisionLimits?.questionTypes ?? []).forEach((type, index) => {
+            if (!OPENAI_DECISIONS_QUESTION_TYPES.has(type)) {
+                issue(["decisionLimits", "questionTypes", index], `decisionLimits.questionTypes cannot list ${type} with dialect openai-decisions: that dialect writes boolean, choice and score questions only`);
+            }
+        });
     }
     else {
         if (mediaKeys !== undefined && mediaAllowance === undefined) {
@@ -778,6 +850,14 @@ const decideCatalogJsonSchema = z
             if (!mediaAllowance.video && mediaKeys.videoKey !== undefined) {
                 issue(["decide", "media", "videoKey"], "decide.media.videoKey must be omitted when decisionLimits.media.video is false");
             }
+            if (mediaAllowance.audio !== undefined &&
+                mediaKeys.audioKey === undefined) {
+                issue(["decide", "media", "audioKey"], "decide.media.audioKey is required when decisionLimits.media.audio is set");
+            }
+            if (mediaAllowance.audio === undefined &&
+                mediaKeys.audioKey !== undefined) {
+                issue(["decide", "media", "audioKey"], "decide.media.audioKey must be omitted when decisionLimits.media.audio is not set");
+            }
         }
     }
     if (decide.baseURLRequired === true &&

package/dist/providers/catalog/xor.json

@@ -10,18 +10,18 @@
         "topModels": ["xor-latest", "xor-omni"],
         "catalog": {
             "xor-latest": {
-                "contextWindow": 250000,
+                "contextWindow": 131072,
                 "maxOutputTokens": 256,
                 "vision": false,
                 "status": "production",
-                "description": "Recommended - XOR text alias; boolean, choice and score questions, 2 to 64 options each"
+                "description": "Recommended - XOR text alias; boolean, choice and score questions over a 131,072-token window. A request with a value question, an image or audio goes to xor-omni when no model was chosen (decide.capableModel)"
             },
             "xor-omni": {
-                "contextWindow": 250000,
+                "contextWindow": 32768,
                 "maxOutputTokens": 256,
                 "vision": true,
                 "status": "production",
-                "description": "XOR multimodal alias; reads images and video alongside the state, for boolean, choice and score questions"
+                "description": "XOR multimodal alias; boolean, choice, score and value questions over a 32,768-token window, plus one image and up to 30 seconds of audio"
             },
             "xor-1.1": {
                 "contextWindow": 250000,
@@ -32,11 +32,24 @@
             }
         }
     },
-    "timeouts": { "decideMs": 5000 },
+    "timeouts": { "decideMs": 5000, "decideMsPerKStateTokens": 150 },
     "decisionLimits": {
-        "maxStateTokens": 200000,
+        "maxStateTokens": 28000,
+        "maxQuestions": 128,
+        "questionTypes": ["boolean", "choice", "score", "value", "choice-unknown"],
+        "maxScoreLevels": 10,
+        "maxChoiceOptions": 64,
         "nonAsciiTokensPerChar": 1,
-        "media": { "maxImages": 8, "video": true, "maxRequestBytes": 8388608 }
+        "models": {
+            "xor-latest": { "maxStateTokens": 120000 },
+            "xor-omni": { "maxStateTokens": 28000 }
+        },
+        "media": {
+            "maxImages": 1,
+            "video": false,
+            "maxRequestBytes": 25000000,
+            "audio": { "maxSeconds": 30 }
+        }
     },
     "decide": {
         "dialect": "system-one",
@@ -46,7 +59,8 @@
         "baseURLShape": "origin",
         "endpoint": "{baseURL}/v1/decisions",
         "model": { "placement": "body" },
-        "media": { "imagesKey": "images", "videoKey": "video" },
+        "media": { "imagesKey": "images", "audioKey": "audio" },
+        "capableModel": { "model": "xor-omni", "questionTypes": ["value"] },
         "requestIdHeaders": [
             "x-request-id",
             "x-typesafe-request-id",
@@ -63,6 +77,12 @@
                     "messagePattern": "(context|prefill|prompt).{0,40}(length|limit|too long|exceed)|(too long|exceeds?).{0,40}(context|tokens)",
                     "kind": "max_tokens_exceeded"
                 },
+                {
+                    "statusMin": 500,
+                    "messagePattern": "backend 4[0-9]{2}",
+                    "kind": "invalid_request",
+                    "retryable": false
+                },
                 { "statusMin": 500, "kind": "server" }
             ],
             "default": "invalid_request",
@@ -91,8 +111,8 @@
     },
     "evidence": {
         "rosterVerified": {
-            "date": "2026-10-18",
-            "method": "POST <origin>/v1/decisions with model xor-latest and a 3-option choice question returned HTTP 200 with served model xor-text-2, through NeuroLink's decide(); a 65-option choice question returned HTTP 422 reading 2..64 options, got 65. A 64-option question returned 200 and 77 options returned 422 on both aliases when probed by hand the same day. A successful xor-omni answer, and any image, audio or video request, were not exercised. The figures contextWindow 250000, maxOutputTokens 256, the 200000-token state window and the media limits are carried over from xor-1.1, not re-measured for the aliases; xor-1.1 is kept only as a retired name"
+            "date": "2026-10-19",
+            "method": "Probed POST <origin>/v1/decisions through a LiteLLM gateway with both aliases. xor-latest read a needle at 120,076 input tokens and refused 200,076 as longer than its 131,072-token context; xor-omni refused past 32,768. 128 questions were answered and 256 refused; a choice took 64 options (63 plus unknown; 64 plus unknown crashed with a 500); a score took 10 levels and 11 were refused by the gateway. value answered int, float and pattern-bound strings on xor-omni, and on xor-latest through the gateway's fallback to xor-omni. One image was read (a second was refused with 422) at a fixed ~600 tokens whatever its resolution. Audio as {data, format} was heard in WAV, FLAC, Ogg, Opus and MP3 (an English speech clip identified at 0.9999 in each) at ~25 tokens a second and cut at 30 seconds, keeping the start; M4A/AAC decoded to nothing. A video, sent as a data URL or as {data, format}, returned the same answer as garbage bytes, so the gateway drops it. The body cap was 24 MiB. A malformed image or audio clip and an over-long state returned 502 wrapping the backend's 400"
         },
         "liveMatrix": null,
         "addedInPR": "https://github.com/juspay/neurolink/pull/1984"

package/dist/providers/configuredDecision.d.ts

@@ -1,4 +1,4 @@
-import type { DecideCatalogJson, DecisionDialect, DecisionError, DecisionPreparedMedia, DecisionState, UnknownRecord } from "../types/index.js";
+import type { DecideCatalogJson, DecisionDialect, DecisionError, DecisionPreparedMedia, DecisionRequest, DecisionState, UnknownRecord } from "../types/index.js";
 import { SystemOneDecisionProvider } from "./systemOneDecision.js";
 /** A `decide` provider described entirely by its catalog entry. */
 export declare class ConfiguredDecisionProvider extends SystemOneDecisionProvider {
@@ -14,6 +14,13 @@ export declare class ConfiguredDecisionProvider extends SystemOneDecisionProvide
     protected missingKeyMessage(): string;
     protected missingConfigMessage(): string | undefined;
     protected decisionEndpoint(): string;
+    /**
+     * `decide.capableModel` in place of the default, for a request that carries
+     * media or one of its question types, when nobody chose a model: the call
+     * names none, the model variable is unset, and the instance is on the
+     * default. A model somebody chose is always the one sent.
+     */
+    protected resolveRequestModel(request: DecisionRequest): string;
     protected decisionHeaders(): Record<string, string>;
     /**
      * The state the dialect is given. Only an entry that sets `stateEncoding`

package/dist/providers/configuredDecision.js

@@ -1,3 +1,4 @@
+import { decisionQuestionTypeOf } from "../utils/decisionLimits.js";
 import { logger } from "../utils/logger.js";
 import { redactUrlForError } from "../utils/logSanitize.js";
 import { getProviderModel } from "../utils/providerConfig.js";
@@ -70,6 +71,27 @@ export class ConfiguredDecisionProvider extends SystemOneDecisionProvider {
     decisionEndpoint() {
         return this.endpoint;
     }
+    /**
+     * `decide.capableModel` in place of the default, for a request that carries
+     * media or one of its question types, when nobody chose a model: the call
+     * names none, the model variable is unset, and the instance is on the
+     * default. A model somebody chose is always the one sent.
+     */
+    resolveRequestModel(request) {
+        const chosen = super.resolveRequestModel(request);
+        const capable = this.entry.decide.capableModel;
+        if (capable === undefined ||
+            request.model !== undefined ||
+            chosen !== this.entry.models.default ||
+            readText(process.env[catalogEnvVar(this.entry, "model")]) !== "") {
+            return chosen;
+        }
+        const carriesMedia = (request.images?.length ?? 0) > 0 ||
+            request.video !== undefined ||
+            request.audio !== undefined;
+        const asksFor = Object.values(request.questions).some((question) => capable.questionTypes.includes(decisionQuestionTypeOf(question)));
+        return carriesMedia || asksFor ? capable.model : chosen;
+    }
     decisionHeaders() {
         const credential = this.entry.decide.authHeader === "x-api-key"
             ? { "x-api-key": this.apiKey }