@juspay/neurolink 12.49.20 → 12.51.0

package/dist/providers/decide/paths.js

@@ -0,0 +1,21 @@
+export function readPath(value, path) {
+    const segments = path.split(".");
+    if (segments.some((segment) => segment === "")) {
+        return undefined;
+    }
+    let current = value;
+    for (const segment of segments) {
+        if (typeof current !== "object" ||
+            current === null ||
+            Array.isArray(current)) {
+            return undefined;
+        }
+        // Own data properties only: no prototype names, and no getter is run.
+        const descriptor = Object.getOwnPropertyDescriptor(current, segment);
+        if (descriptor === undefined || !("value" in descriptor)) {
+            return undefined;
+        }
+        current = descriptor.value;
+    }
+    return current;
+}

package/dist/providers/decide/template.d.ts

@@ -0,0 +1,3 @@
+export declare function fillEndpointTemplate(template: string, parts: {
+    baseURL: string;
+}): string;

package/dist/providers/decide/template.js

@@ -0,0 +1,13 @@
+const PLACEHOLDER = /\{([^{}]*)\}/g;
+export function fillEndpointTemplate(template, parts) {
+    if (/[{}]/.test(template.replace(PLACEHOLDER, ""))) {
+        throw new Error(`Decide endpoint template "${template}" has an unmatched brace; write placeholders as {baseURL}.`);
+    }
+    // A function replacer, so "$&" or "$1" inside a base URL is inserted as text.
+    return template.replace(PLACEHOLDER, (_match, name) => {
+        if (name !== "baseURL") {
+            throw new Error(`Decide endpoint template "${template}" uses an unknown placeholder {${name}}; the only one supported is {baseURL}.`);
+        }
+        return parts.baseURL;
+    });
+}

package/dist/types/conversation.d.ts

@@ -295,6 +295,8 @@ export type ToolResultData = {
     /** Error message if execution failed */
     error?: string;
 };
+/** How a conversation turn was entered. */
+export type ConversationModality = "text" | "voice";
 /** Metadata associated with a ChatMessage. */
 export type ChatMessageMetadata = {
     /** Is this a summary message? */
@@ -328,6 +330,22 @@ export type ChatMessageMetadata = {
     thinkingExpanded?: boolean;
     /** Step index for reconstructing parallel vs sequential tool calls */
     stepIndex?: number;
+    /**
+     * How the message was entered: typed (`"text"`) or spoken (`"voice"`).
+     * Absent on every message stored before this field existed, and on typed
+     * turns that did not set it; readers treat absence as `"text"`.
+     */
+    modality?: ConversationModality;
+    /**
+     * Stable id of the voice turn this message belongs to. The stored message
+     * ids are `${turnId}:user` and `${turnId}:assistant`, which is what makes
+     * recording the same turn twice a no-op.
+     */
+    turnId?: string;
+    /** True when the message was cut off (barge-in, or the call ended) before it finished. */
+    interrupted?: boolean;
+    /** True when the call ended while this turn was still open. */
+    callEnded?: boolean;
     /**
      * Head/tail preview of a large tool output.
      * Only present on tool_result messages where the output exceeded truncation limits.
@@ -515,6 +533,97 @@ export type StoreConversationTurnOptions = {
      * for the canonical partition-and-reanchor pattern).
      */
     skillMessages?: ChatMessage[];
+    /**
+     * How the turn was entered. Stamped on both stored messages' metadata when
+     * set; a typed turn that leaves it unset is stored exactly as before.
+     */
+    modality?: ConversationModality;
+    /**
+     * Idempotency key for the turn. When set, the user and assistant messages
+     * are stored with the ids `${turnId}:user` and `${turnId}:assistant`, and a
+     * turn whose assistant id is already in the session's recent messages is
+     * skipped, so a retry after an ambiguous failure cannot duplicate it.
+     */
+    turnId?: string;
+    /** The assistant message was cut off before it finished. */
+    interrupted?: boolean;
+    /** The call ended while this turn was open. */
+    callEnded?: boolean;
+};
+/** Id and metadata a memory manager gives one stored message of a turn. */
+export type TurnMessageIdentity = {
+    /** Stable id for the message, or undefined to let the manager mint a random one. */
+    id: string | undefined;
+    /** Metadata to set on the message, or undefined when there is none to add. */
+    metadata: ChatMessageMetadata | undefined;
+};
+/**
+ * Input to `NeuroLink.recordConversationTurn`: one finished turn from a
+ * transport that does not go through generate()/stream() (a voice call),
+ * written to the same conversation memory typed chat uses.
+ */
+export type RecordConversationTurnOptions = {
+    /** Conversation to append to. Must not contain ":" (it is part of the stored key). */
+    sessionId: string;
+    /**
+     * Owner of the conversation. Required: typed chat falls back to a shared
+     * placeholder user when it is missing, and a spoken turn must never land
+     * there.
+     */
+    userId: string;
+    /** What the person said. Must be non-blank. */
+    userMessage: string;
+    /** What the assistant said (or the closing note for a cut-off turn). Must be non-blank unless `events` carries tool activity. */
+    aiResponse: string;
+    /** How the turn was entered. Stamped on both stored messages. */
+    modality?: ConversationModality;
+    /** Idempotency key. Recording the same `turnId` again is a no-op. */
+    turnId?: string;
+    /** The assistant message was cut off before it finished. */
+    interrupted?: boolean;
+    /** The call ended while this turn was open. */
+    callEnded?: boolean;
+    /** Tool and UI events of the turn, in the shape typed chat stores. */
+    events?: StreamEventSequence[];
+    /** When the turn started. */
+    startTime?: Date;
+    /** Provider and model that answered, for token-threshold selection. */
+    provider?: string;
+    model?: string;
+    /** Observability request identifier. */
+    requestId?: string;
+    /** Per-call override of summarization, as on generate()/stream(). */
+    enableSummarization?: boolean;
+    /** Per-call long-term memory switches, as on generate()/stream(). */
+    memory?: {
+        enabled?: boolean;
+        write?: boolean;
+    };
+};
+/** Outcome of `NeuroLink.recordConversationTurn`. */
+export type RecordConversationTurnResult = {
+    /** False when the turn was deliberately not stored; see `reason`. */
+    recorded: boolean;
+    /** Why nothing was stored. Only set when `recorded` is false. */
+    reason?: "empty-user-message" | "empty-response";
+    sessionId: string;
+    userId: string;
+    turnId?: string;
+};
+/** Options for `NeuroLink.getConversationSeed`. */
+export type ConversationSeedOptions = {
+    sessionId: string;
+    userId: string;
+    /** Most recent user/assistant exchanges to return (default 10). */
+    maxTurns?: number;
+    /** Cap on the total characters returned; the oldest messages are dropped first (default 6000). */
+    maxChars?: number;
+};
+/** One message of a conversation seed. */
+export type ConversationSeedMessage = {
+    role: "user" | "assistant";
+    content: string;
+    modality?: ConversationModality;
 };
 /**
  * Lightweight session metadata for efficient session listing

package/dist/types/livekit.d.ts

@@ -11,6 +11,7 @@
  */
 import type { NeuroLinkEvents, TypedEventEmitter } from "./common.js";
 import type { StreamOptions, StreamResult } from "./stream.js";
+import type { ConversationSeedMessage, ConversationSeedOptions, RecordConversationTurnOptions, RecordConversationTurnResult, StreamEventSequence } from "./conversation.js";
 /**
  * Minimal structural shape of the NeuroLink instance the brain depends on.
  *
@@ -57,6 +58,18 @@ export type LiveKitBrainTurn = {
     conversationId: string;
     /** Cancellation signal; aborting stops the in-flight LLM and tool calls. */
     signal?: AbortSignal;
+    /**
+     * User the conversation belongs to for THIS call, overriding the brain's
+     * static `userId`. A call whose memory binding verified passes the bound
+     * user here so the turn is stored under the person on the call.
+     */
+    userId?: string;
+    /**
+     * Run this turn without reading or writing conversation memory. Used for the
+     * agent's opening greeting on a bound call, which is an instruction to the
+     * agent and must not land in the person's conversation as something they said.
+     */
+    ephemeral?: boolean;
 };
 /**
  * The brain's public surface: stream the assistant reply as text deltas.
@@ -155,6 +168,15 @@ export type LiveKitVoiceAgentConfig = {
     conversationIdPrefix?: string;
     /** Optional user id recorded alongside memory. */
     userId?: string;
+    /**
+     * Per-call memory identity. When set, the call's room metadata is checked
+     * for a signed memory binding (see `signVoiceMemoryBinding`); a binding that
+     * verifies replaces `userId` and the room-derived conversation id for that
+     * call, and the turns the brain streams are stamped as voice. A call whose
+     * binding is missing or does not verify runs with the static `userId` and
+     * the room-derived conversation id exactly as before.
+     */
+    memory?: LiveKitVoiceAgentMemoryConfig;
     greeting?: string;
     /** Silero VAD tuning (stricter = ignores background noise). */
     vad?: LiveKitVadConfig;
@@ -170,6 +192,11 @@ export type LiveKitVoiceAgentConfig = {
      */
     events?: LiveKitEventBridgeConfig;
 };
+/** Per-call memory identity for the cascaded voice agent. */
+export type LiveKitVoiceAgentMemoryConfig = {
+    /** Shared secret the binding is signed with. Defaults to `NEUROLINK_VOICE_MEMORY_SECRET`. */
+    secret?: string;
+};
 /** Options for `startVoiceAgentWorker` — launches the LiveKit Agents worker. */
 export type LiveKitWorkerLaunchOptions = {
     /**
@@ -208,7 +235,7 @@ export type LiveKitTokenRequest = {
     ttlSeconds?: number;
 };
 /** Discriminant tags for outbound voice events. */
-export type LiveKitVoiceEventType = "user-text" | "text" | "tool-start" | "tool-result" | "status" | "hitl-prompt" | "done";
+export type LiveKitVoiceEventType = "user-text" | "text" | "tool-start" | "tool-result" | "status" | "hitl-prompt" | "done" | "memory";
 /**
  * A user STT transcript for display. Interim partials stream with
  * `final: false`; the end-of-utterance result has `final: true`. The client
@@ -277,15 +304,32 @@ export type LiveKitVoiceHitlPromptEvent = {
         allowModification?: boolean;
     };
 };
-/** The current turn has finished. */
+/**
+ * The current turn has finished. `turnId` is set when the call's turns are
+ * being recorded: it is the id the stored messages carry (`${turnId}:user`,
+ * `${turnId}:assistant`), so a client can merge its live copy with history.
+ */
 export type LiveKitVoiceDoneEvent = {
     type: "done";
     data: {
         reason?: string;
+        turnId?: string;
+    };
+};
+/**
+ * Something about how the call is being remembered. `not-saved` means a turn
+ * could not be stored after retries; the call carries on, and a client may
+ * show a quiet note on that turn.
+ */
+export type LiveKitVoiceMemoryEvent = {
+    type: "memory";
+    data: {
+        state: "not-saved";
+        turnId: string;
     };
 };
 /** Discriminated union of all outbound voice events (before enveloping). */
-export type LiveKitVoiceEvent = LiveKitVoiceUserTextEvent | LiveKitVoiceTextEvent | LiveKitVoiceToolStartEvent | LiveKitVoiceToolResultEvent | LiveKitVoiceStatusEvent | LiveKitVoiceHitlPromptEvent | LiveKitVoiceDoneEvent;
+export type LiveKitVoiceEvent = LiveKitVoiceUserTextEvent | LiveKitVoiceTextEvent | LiveKitVoiceToolStartEvent | LiveKitVoiceToolResultEvent | LiveKitVoiceStatusEvent | LiveKitVoiceHitlPromptEvent | LiveKitVoiceDoneEvent | LiveKitVoiceMemoryEvent;
 /**
  * Wire format published to the browser: a `LiveKitVoiceEvent` plus a monotonic
  * sequence number and a timestamp so the client can order and de-duplicate.
@@ -442,6 +486,14 @@ export type RealtimeVoiceAgentConfig = {
         enabled?: boolean;
         mcpUrl?: string;
     };
+    /**
+     * Persist the call's turns to conversation memory and seed the model with
+     * the conversation's recent history. Off unless set. Takes effect only for
+     * a call whose room metadata carries a verified memory binding; otherwise
+     * the call runs exactly as it would without this option, and nothing is
+     * stored.
+     */
+    memory?: RealtimeVoiceMemoryConfig;
     /** Data-channel topic for outbound events (default "ai-events"). */
     eventsTopic?: string;
     /** Data-channel topic for inbound control messages (default "ai-control"). */
@@ -455,6 +507,186 @@ export type RealtimeVoiceAgentConfig = {
      */
     onLog?: (entry: RealtimeVoiceLogEntry, ctx: RealtimeVoiceLogContext) => void;
 };
+/**
+ * The part of a NeuroLink instance a voice agent needs to persist and seed a
+ * call's turns. The real `NeuroLink` instance satisfies it.
+ */
+export type LiveKitMemoryNeuroLink = {
+    recordConversationTurn: (options: RecordConversationTurnOptions) => Promise<RecordConversationTurnResult>;
+    getConversationSeed?: (options: ConversationSeedOptions) => Promise<ConversationSeedMessage[]>;
+};
+/** Memory options for `defineRealtimeVoiceAgent`. */
+export type RealtimeVoiceMemoryConfig = {
+    /**
+     * Builds the NeuroLink instance, with conversation memory configured the
+     * same way the web app's is (same store, key prefix and lifetime), inside
+     * the job process. Called once per call, only after a binding verified.
+     */
+    createNeuroLink: () => LiveKitMemoryNeuroLink | Promise<LiveKitMemoryNeuroLink>;
+    /** Shared secret the binding is signed with. Defaults to `NEUROLINK_VOICE_MEMORY_SECRET`. */
+    secret?: string;
+    /** How long a closed reply waits for its user transcript before being dropped (default 1500 ms). */
+    settleMs?: number;
+    /** How long shutdown waits for pending writes (default 800 ms; the worker is reaped about 1 s after shutdown starts). */
+    flushBudgetMs?: number;
+    /** History seeding at call start. `false` turns it off. Defaults: 10 exchanges, 6000 characters. */
+    seed?: false | {
+        maxTurns?: number;
+        maxChars?: number;
+    };
+};
+/** One finished exchange of a voice call, as handed to the writer. */
+export type VoiceTurnRecord = {
+    /** The assistant message id, or a synthetic id for a turn cut off by shutdown or barge-in. */
+    turnId: string;
+    userMessage: string;
+    aiResponse: string;
+    /** The reply was cut off. */
+    interrupted: boolean;
+    /** The call ended while this exchange was open. */
+    callEnded: boolean;
+    startTime: Date;
+    /** Tool events of the exchange, in the shape typed chat stores. */
+    events?: StreamEventSequence[];
+};
+/** A recorder's current exchange: what has been heard and said since the last reply closed. */
+export type VoiceTurnOpenExchange = {
+    userSegments: string[];
+    assistantPartial: string;
+    assistantStarted: boolean;
+    events: StreamEventSequence[];
+    startedAt: number;
+};
+/** A reply that finished before its user transcript arrived, waiting out the settle window. */
+export type VoiceTurnClosedExchange = {
+    turnId: string;
+    assistantText: string;
+    interrupted: boolean;
+    events: StreamEventSequence[];
+    startedAt: number;
+    timer: ReturnType<typeof setTimeout>;
+};
+/** The recorder returned by `createVoiceTurnRecorder`. */
+export type VoiceTurnRecorder = {
+    /** A final transcript segment of what the user said. Interim segments are ignored. */
+    userTranscript: (segment: {
+        text: string;
+        final: boolean;
+    }) => void;
+    /** A piece of the assistant's reply as it is produced. */
+    assistantDelta: (delta: string) => void;
+    /** The assistant's message was added: the reply is complete (or was cut off). */
+    assistantMessage: (message: {
+        id: string;
+        text: string;
+        interrupted?: boolean;
+    }) => void;
+    /**
+     * The next assistant-only reply is the agent's opening greeting, not an
+     * answer: it is skipped and never paired with the user's first question.
+     * Call it once, when the agent is about to speak its greeting.
+     */
+    expectGreeting: () => void;
+    /** A tool call started in the current exchange. */
+    toolStart: (call: {
+        name: string;
+        args?: unknown;
+    }) => void;
+    /** A tool call finished in the current exchange. */
+    toolEnd: (call: {
+        name: string;
+        result?: unknown;
+        error?: string;
+    }) => void;
+    /**
+     * Shutdown: close what is open, then wait for pending writes — but no longer
+     * than `budgetMs`. Safe to call more than once.
+     */
+    flush: (budgetMs: number) => Promise<VoiceTurnRecorderStats>;
+    /** Counters, for logs and tests. */
+    stats: () => VoiceTurnRecorderStats;
+};
+/** What a realtime call needs from its memory once its binding verified. */
+export type RealtimeCallMemory = {
+    recorder: VoiceTurnRecorder;
+    /** Recent history to seed the model with; empty when there is none or seeding is off. */
+    seed: ConversationSeedMessage[];
+    /** Late-bound publisher, set once the event bridge exists, for the not-saved notice. */
+    setPublisher: (publish: (type: LiveKitVoiceEventType, data: Record<string, unknown>) => void) => void;
+};
+/** Counters of a recorder, for logs and tests. */
+export type VoiceTurnRecorderStats = {
+    written: number;
+    failed: number;
+    dropped: number;
+    pending: number;
+};
+/** Options for `createVoiceTurnRecorder`. */
+export type VoiceTurnRecorderOptions = {
+    /** Stores one exchange. Throwing means "not stored"; the recorder retries. */
+    write: (record: VoiceTurnRecord) => Promise<void>;
+    /** A write succeeded. */
+    onRecorded?: (record: VoiceTurnRecord) => void;
+    /** A write failed after every retry. The call continues. */
+    onPermanentFailure?: (record: VoiceTurnRecord, error: unknown) => void;
+    /** How long a closed reply waits for its user transcript (default 1500 ms). */
+    settleMs?: number;
+    /** Retries after the first failed attempt (default 2). */
+    retries?: number;
+    /** Base delay between attempts; grows linearly (default 250 ms). */
+    retryDelayMs?: number;
+    /** Appended to a reply that was cut off by the call ending. */
+    callEndedNote?: string;
+    /** Stored as the reply when the call ended with the user's turn unanswered. */
+    noReplyNote?: string;
+    /** Prefix of synthetic turn ids (default: a random id). */
+    callId?: string;
+    /** Clock, for tests. */
+    now?: () => number;
+};
+/** A signed memory binding as it appears in room metadata under `memory`. */
+export type VoiceMemoryBinding = {
+    userId: string;
+    sessionId: string;
+    /** Expiry, Unix seconds. */
+    exp: number;
+    /** base64url(HMAC-SHA256) over the room name, userId, sessionId and exp. */
+    sig: string;
+};
+/** Input to `signVoiceMemoryBinding`. */
+export type VoiceMemoryBindingInput = {
+    /** The LiveKit room name the binding is for. */
+    room: string;
+    userId: string;
+    sessionId: string;
+    secret: string;
+    /** Lifetime in seconds (default 7200, at most 86400). */
+    ttlSeconds?: number;
+    /** Current time, Unix seconds (default: now). */
+    now?: number;
+};
+/** Input to `verifyVoiceMemoryBinding`. */
+export type VoiceMemoryBindingVerifyInput = {
+    /** The room's base64(JSON) metadata string. */
+    roomMetadata: string | undefined;
+    /** The room name as LiveKit reports it for the joined room. */
+    roomName: string | undefined;
+    secret: string | undefined;
+    /** Current time, Unix seconds (default: now). */
+    now?: number;
+};
+/** Why a binding was not accepted. */
+export type VoiceMemoryBindingRejection = "no-secret" | "no-binding" | "malformed" | "bad-signature" | "expired" | "expiry-too-far";
+/** Result of `verifyVoiceMemoryBinding`. */
+export type VoiceMemoryBindingVerdict = {
+    ok: true;
+    userId: string;
+    sessionId: string;
+    exp: number;
+} | {
+    ok: false;
+    reason: VoiceMemoryBindingRejection;
+};
 /** Auth token + base64 MCP execution context decoded from a room's metadata. */
 export type LiveKitRoomCallContext = {
     /** Lighthouse access JWT used as `x-auth-token` to the MCP server. */
@@ -509,4 +741,15 @@ export type BuildRealtimeMcpToolsParams = {
      * model can talk about.
      */
     toolTimeoutMs?: number;
+    /** Observes a tool call as it starts. A throwing observer never affects the call. */
+    onToolStart?: (call: {
+        name: string;
+        args: Record<string, unknown>;
+    }) => void;
+    /** Observes a tool call as it ends, with its result as text or its error. */
+    onToolEnd?: (call: {
+        name: string;
+        result?: string;
+        error?: string;
+    }) => void;
 };

package/dist/types/providerCatalog.d.ts

@@ -4,6 +4,7 @@
  * src/lib/providers/catalog/<id>.json file per provider is the single
  * source of truth; see docs/superpowers/plans/2026-08-28-provider-json-catalog-spec.md.
  */
+import type { DecideDialectName, DecisionErrorKind, DecisionLimits } from "./decision.js";
 export type CatalogPricingPerMTok = {
     input: number;
     output: number;
@@ -196,7 +197,7 @@ export type ProviderCatalogJson = {
      */
     credentialsKey?: string;
     aliases: string[];
-    tier: 2;
+    tier?: 2;
     wire: CatalogWire;
     models: {
         default: string;
@@ -254,3 +255,75 @@ export type ProviderCatalogJson = {
     setup: CatalogSetup;
     evidence: CatalogEvidence;
 };
+export type DecideFormatterName = "fastapi-validation";
+export type DecideBaseURLTransform = "strip-trailing-slash";
+export type DecideErrorRule = {
+    /** One status or a list of them. A rule sets this or `statusMin`, never both. */
+    status?: number | readonly number[];
+    /** Matches every status at or above this value. */
+    statusMin?: number;
+    kind: DecisionErrorKind;
+    retryable?: boolean;
+};
+export type DecideBodyShape = {
+    /** Dot path into the parsed error body. */
+    path: string;
+    isArray?: boolean;
+    formatter?: DecideFormatterName;
+    kind?: DecisionErrorKind;
+    retryable?: boolean;
+};
+export type DecideErrors = {
+    rules: readonly DecideErrorRule[];
+    default: DecisionErrorKind;
+    bodyShapes: readonly DecideBodyShape[];
+};
+export type DecideModel = {
+    placement: "body" | "body-unless-alias";
+    /** With `body-unless-alias`: the model name that means "send no model". */
+    alias?: string;
+};
+export type DecideBlock = {
+    dialect: DecideDialectName;
+    /** Decision order, lower runs first; unique across decide providers. */
+    order: number;
+    /** True when the provider has no built-in endpoint and the caller must supply one. */
+    baseURLRequired?: boolean;
+    baseURLNormalize?: readonly DecideBaseURLTransform[];
+    /** Request URL; `{baseURL}` is the only placeholder. */
+    endpoint: string;
+    model: DecideModel;
+    requestIdHeaders?: readonly string[];
+    response?: {
+        modelPaths?: readonly string[];
+    };
+    errors: DecideErrors;
+};
+export type DecideCatalogJson = {
+    $schema?: string;
+    id: string;
+    displayName: string;
+    enumTypeName?: string;
+    credentialsKey?: string;
+    aliases: string[];
+    inferenceKinds: readonly ["decide"];
+    tier?: 2;
+    wire: CatalogWire;
+    models: {
+        default: string;
+        catalog: Record<string, CatalogModelSpec>;
+        topModels?: string[];
+        testModel?: string;
+    };
+    timeouts?: {
+        decideMs?: number;
+    };
+    decisionLimits?: DecisionLimits;
+    decide: DecideBlock;
+    setup: CatalogSetup;
+    evidence: CatalogEvidence;
+};
+export type CatalogEntry = ProviderCatalogJson | DecideCatalogJson;
+export type CatalogKeySource = Pick<CatalogEntry, "id" | "credentialsKey">;
+export type CatalogEnvSource = Pick<CatalogEntry, "id" | "wire">;
+export type CatalogSetupSource = Pick<CatalogEntry, "id" | "displayName" | "wire" | "setup">;

package/dist/types/providerCatalog.generated.d.ts

@@ -1,2 +1,2 @@
-export type CatalogProviderName = "a2agent" | "abliteration" | "above-dev" | "ai21" | "aiand" | "aihubmix" | "aionlabs" | "ambient" | "api-route" | "arcee" | "atlas-cloud" | "avian-io" | "baidu-qianfan" | "baseten" | "bee-heossi" | "byteplus-modelark" | "bytez" | "cerebras" | "charm-hyper" | "chutes" | "cloudflare" | "crusoe" | "dashscope" | "deepinfra" | "deepseek" | "empiriolabs" | "featherless-ai" | "fireworks" | "friendli" | "gmicloud" | "gradientai" | "groq" | "hetzner-inference" | "huggingface" | "inception-labs" | "inceptron" | "inco" | "inference-net" | "io-intelligence" | "koscompute" | "lemonfox-ai" | "lilac" | "llmtech" | "mancer" | "meta-model-api" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshot-ai" | "morph" | "naver-clova-studio" | "nebius" | "neuralwatt" | "novita" | "ovhcloud" | "parasail" | "pareto-inference" | "perplexity" | "poolside" | "prime-intellect" | "reka" | "sakana-ai" | "sambanova" | "sarvam" | "scaleway" | "siliconflow" | "stackit" | "stepfun" | "subconscious" | "synthetic" | "telnyx" | "thinking-machines" | "tinfoil" | "together-ai" | "umans-ai" | "upstage" | "venice-ai" | "vispark" | "vultr-inference" | "wafer" | "wandb-inference" | "xai" | "z-ai";
-export type CatalogCredentialKey = "a2agent" | "abliteration" | "aboveDev" | "ai21" | "aiand" | "aihubmix" | "aionlabs" | "ambient" | "apiRoute" | "arcee" | "atlasCloud" | "avianIo" | "baiduQianfan" | "baseten" | "beeHeossi" | "byteplusModelark" | "bytez" | "cerebras" | "charmHyper" | "chutes" | "cloudflare" | "crusoe" | "dashscope" | "deepinfra" | "deepseek" | "empiriolabs" | "featherlessAi" | "fireworks" | "friendli" | "gmicloud" | "gradientai" | "groq" | "hetznerInference" | "huggingFace" | "inceptionLabs" | "inceptron" | "inco" | "inferenceNet" | "ioIntelligence" | "koscompute" | "lemonfoxAi" | "lilac" | "llmtech" | "mancer" | "metaModelApi" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshotAi" | "morph" | "naverClovaStudio" | "nebius" | "neuralwatt" | "novita" | "ovhcloud" | "parasail" | "paretoInference" | "perplexity" | "poolside" | "primeIntellect" | "reka" | "sakanaAi" | "sambanova" | "sarvam" | "scaleway" | "siliconflow" | "stackit" | "stepfun" | "subconscious" | "synthetic" | "telnyx" | "thinkingMachines" | "tinfoil" | "together" | "umansAi" | "upstage" | "veniceAi" | "vispark" | "vultrInference" | "wafer" | "wandbInference" | "xai" | "zAi";
+export type CatalogProviderName = "a2agent" | "abliteration" | "above-dev" | "ai21" | "aiand" | "aihubmix" | "aionlabs" | "ambient" | "api-route" | "arcee" | "atlas-cloud" | "avian-io" | "baidu-qianfan" | "baseten" | "bee-heossi" | "byteplus-modelark" | "bytez" | "cerebras" | "charm-hyper" | "chutes" | "cloudflare" | "crusoe" | "dashscope" | "deepinfra" | "deepseek" | "empiriolabs" | "featherless-ai" | "fireworks" | "friendli" | "gmicloud" | "gradientai" | "groq" | "hetzner-inference" | "huggingface" | "inception-labs" | "inceptron" | "inco" | "inference-net" | "io-intelligence" | "koscompute" | "laya" | "lemonfox-ai" | "lilac" | "llmtech" | "mancer" | "meta-model-api" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshot-ai" | "morph" | "naver-clova-studio" | "nebius" | "neuralwatt" | "novita" | "ovhcloud" | "parasail" | "pareto-inference" | "perplexity" | "poolside" | "prime-intellect" | "reka" | "sakana-ai" | "sambanova" | "sarvam" | "scaleway" | "siliconflow" | "stackit" | "stepfun" | "subconscious" | "synthetic" | "telnyx" | "thinking-machines" | "tinfoil" | "together-ai" | "umans-ai" | "upstage" | "venice-ai" | "vispark" | "vultr-inference" | "wafer" | "wandb-inference" | "xai" | "z-ai";
+export type CatalogCredentialKey = "a2agent" | "abliteration" | "aboveDev" | "ai21" | "aiand" | "aihubmix" | "aionlabs" | "ambient" | "apiRoute" | "arcee" | "atlasCloud" | "avianIo" | "baiduQianfan" | "baseten" | "beeHeossi" | "byteplusModelark" | "bytez" | "cerebras" | "charmHyper" | "chutes" | "cloudflare" | "crusoe" | "dashscope" | "deepinfra" | "deepseek" | "empiriolabs" | "featherlessAi" | "fireworks" | "friendli" | "gmicloud" | "gradientai" | "groq" | "hetznerInference" | "huggingFace" | "inceptionLabs" | "inceptron" | "inco" | "inferenceNet" | "ioIntelligence" | "koscompute" | "laya" | "lemonfoxAi" | "lilac" | "llmtech" | "mancer" | "metaModelApi" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshotAi" | "morph" | "naverClovaStudio" | "nebius" | "neuralwatt" | "novita" | "ovhcloud" | "parasail" | "paretoInference" | "perplexity" | "poolside" | "primeIntellect" | "reka" | "sakanaAi" | "sambanova" | "sarvam" | "scaleway" | "siliconflow" | "stackit" | "stepfun" | "subconscious" | "synthetic" | "telnyx" | "thinkingMachines" | "tinfoil" | "together" | "umansAi" | "upstage" | "veniceAi" | "vispark" | "vultrInference" | "wafer" | "wandbInference" | "xai" | "zAi";

package/dist/types/providers.d.ts

@@ -353,6 +353,10 @@ export type NeurolinkCredentials = {
         apiKey?: string;
         baseURL?: string;
     };
+    laya?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     lemonfoxAi?: {
         apiKey?: string;
         baseURL?: string;
@@ -585,15 +589,6 @@ export type NeurolinkCredentials = {
          */
         gatewayURL?: string;
     };
-    /**
-     * Laya (Convai, open weights) — the `decide` inference type. There is no
-     * built-in endpoint: `baseURL` (or LAYA_BASE_URL) is required, pointing at a
-     * Laya server or a proxy route to one; requests go to `<baseURL>/predict`.
-     */
-    laya?: {
-        apiKey?: string;
-        baseURL?: string;
-    };
     /**
      * XOR (Juspay, open weights) — the `decide` inference type. There is no
      * built-in endpoint: `baseURL` (or XOR_BASE_URL) is required, pointing at the
@@ -2447,6 +2442,14 @@ export type ProviderDescriptor = {
      * checks to its own server (TypeSafe). See {@link DecisionLimits}.
      */
     decisionLimits?: DecisionLimits;
+    /**
+     * Precedence among `decide` providers: lower runs first and no two decide
+     * descriptors share a value. resolveDefaultDecisionProvider() returns the
+     * first configured provider in this order, so changing a number changes which
+     * model routes for every host that configures several. Required on every
+     * descriptor that serves `decide`; checked when the descriptors load.
+     */
+    decideOrder?: number;
     /**
      * Ascending priority (1 = tried first) in the auto-select fallback chain
      * used by getBestProvider(). Undefined = not part of the auto-select chain.

package/dist/types/sdk.d.ts

@@ -12,7 +12,7 @@ export type { AnalyticsData, TokenUsage } from "./analytics.js";
 export type { BaseCommandArgs, CommandResult, GenerateCommandArgs, CliGenerateResult, MCPCommandArgs, ModelsCommandArgs, CliStreamChunk, } from "./cli.js";
 export type { ErrorInfo, FunctionParameters, JsonArray, JsonObject, JsonValue, ProcessResult, Result, TestFunction, TestResult, Unknown, UnknownArray, UnknownRecord, } from "./common.js";
 export type { Content, ImageContent, MultimodalMessage, ProcessedImage, ProviderImageFormat, ProviderMultimodalPayload, TextContent, VisionCapability, } from "./content.js";
-export type { ChatMessage, ChatMessageMetadata, ToolResultData, ConversationMemoryConfig, ConversationMemoryError, ConversationMemoryEvents, MessageContent, MultimodalChatMessage, RedisConversationObject, RedisStorageConfig, SessionIdentifier, SessionMemory, SessionMetadata, StoreConversationTurnOptions, } from "./conversation.js";
+export type { ChatMessage, ChatMessageMetadata, ToolResultData, ConversationMemoryConfig, ConversationMemoryError, ConversationMemoryEvents, ConversationModality, ConversationSeedMessage, ConversationSeedOptions, MessageContent, MultimodalChatMessage, RecordConversationTurnOptions, RecordConversationTurnResult, RedisConversationObject, RedisStorageConfig, SessionIdentifier, SessionMemory, SessionMetadata, StoreConversationTurnOptions, } from "./conversation.js";
 export type { DomainConfig, DomainConfigOptions, DomainEvaluationCriteria, DomainTemplate, DomainType, DomainValidationRule, } from "./domain.js";
 export type { EnhancedEvaluationResult, EvaluationContext, EvaluationCriteria, EvaluationData, EvaluationRequest, } from "./evaluation.js";
 export type { ExternalMCPConfigValidation, ExternalMCPManagerConfig, ExternalMCPOperationResult, ExternalMCPServerEvents, ExternalMCPServerHealth, ExternalMCPServerInstance, ExternalMCPServerStatus, ExternalMCPToolContext, ExternalMCPToolInfo, ExternalMCPToolResult, } from "./externalMcp.js";

package/dist/types/stream.d.ts

@@ -9,7 +9,7 @@ import type { TokenUsage } from "./analytics.js";
 import type { JsonValue, UnknownRecord } from "./common.js";
 import type { TerminalAgentModeOption, TerminalAgentModeVersion } from "./agentMode.js";
 import type { Content, ImageWithAltText } from "./content.js";
-import type { ChatMessage, ToolReplayMode } from "./conversation.js";
+import type { ChatMessage, ConversationModality, ToolReplayMode } from "./conversation.js";
 import type { StreamNoOutputSentinel } from "./noOutputSentinel.js";
 import type { GenerateStopReason, TTSMetadata, ToolExecutionCaptureOptions } from "./generate.js";
 import type { MemoryCallOptions } from "./memory.js";
@@ -666,6 +666,13 @@ export type StreamOptions = {
     workflow?: string;
     workflowConfig?: WorkflowConfig;
     enableSummarization?: boolean;
+    /**
+     * How this turn was entered. Stamped on the stored messages' metadata so
+     * history readers can tell a spoken turn from a typed one. Leave it unset
+     * for typed input: the stored turn is then identical to what it was before
+     * the field existed.
+     */
+    modality?: ConversationModality;
     /**
      * Maximum cumulative cost (USD) for this session.
      * Once the session spend reaches this limit, subsequent stream() calls