@juspay/neurolink 12.56.1 → 12.58.0

package/dist/telemetry/attributes.js

@@ -140,8 +140,24 @@ export const LANGFUSE_ATTR = {
 };
 /** Default ceiling for serialized span attribute values. */
 export const SPAN_ATTRIBUTE_MAX_CHARS = 40_000;
+/** Smallest ceiling accepted, so the system and question openings still fit. */
+const SPAN_ATTRIBUTE_MIN_CHARS = 10_000;
+/**
+ * The ceiling in force: `NEUROLINK_SPAN_ATTRIBUTE_MAX_CHARS` when it is a
+ * positive whole number (raised to the minimum if lower), otherwise the
+ * default. Read on every call, so a change applies to the next span.
+ */
+export function spanAttributeMaxChars() {
+    const configured = Number(process.env.NEUROLINK_SPAN_ATTRIBUTE_MAX_CHARS?.trim());
+    if (!Number.isSafeInteger(configured) || configured <= 0) {
+        return SPAN_ATTRIBUTE_MAX_CHARS;
+    }
+    return Math.max(SPAN_ATTRIBUTE_MIN_CHARS, configured);
+}
 /** Opening of the system prompt a model-request attribute always keeps. */
 const SYSTEM_PROMPT_FLOOR_CHARS = 2_000;
+/** Room the current question always gets in a model-request attribute. */
+const QUESTION_FLOOR_CHARS = 4_000;
 /** Room for the "...[truncated N chars]" note, kept outside that opening. */
 const TRUNCATION_NOTE_ALLOWANCE_CHARS = 48;
 const MIN_SHRINKABLE_STRING_CHARS = 64;
@@ -288,7 +304,7 @@ function shrinkLongestString(root, excess) {
  * A structured value over the cap is shortened by trimming its longest
  * strings, so the attribute stays valid JSON instead of ending mid-token.
  */
-export function spanContentAttribute(value, maxChars = SPAN_ATTRIBUTE_MAX_CHARS) {
+export function spanContentAttribute(value, maxChars = spanAttributeMaxChars()) {
     const redacted = redactSafely(value);
     if (redacted === null || typeof redacted !== "object") {
         return spanJsonAttribute(redacted, maxChars);
@@ -314,7 +330,7 @@ export function spanContentAttribute(value, maxChars = SPAN_ATTRIBUTE_MAX_CHARS)
  * on a single span. Strings pass through unserialized; everything else is
  * JSON-stringified with a String() fallback for circular structures.
  */
-export function spanJsonAttribute(value, maxChars = SPAN_ATTRIBUTE_MAX_CHARS) {
+export function spanJsonAttribute(value, maxChars = spanAttributeMaxChars()) {
     let serialized;
     try {
         serialized =
@@ -332,17 +348,72 @@ export function spanJsonAttribute(value, maxChars = SPAN_ATTRIBUTE_MAX_CHARS) {
     }
     return serialized;
 }
+/** The content blocks of an Anthropic-shaped message; none for plain text. */
+function messageBlocks(message) {
+    if (message === null || typeof message !== "object") {
+        return [];
+    }
+    return "content" in message && Array.isArray(message.content)
+        ? message.content
+        : [];
+}
+/** One field of a content block, or undefined when the block lacks it. */
+function blockField(block, field) {
+    return block !== null && typeof block === "object" && field in block
+        ? Reflect.get(block, field)
+        : undefined;
+}
+/** Whether a message is the user's, as opposed to the assistant's. */
+function isUserMessage(message) {
+    return (message !== null &&
+        typeof message === "object" &&
+        "role" in message &&
+        message.role === "user");
+}
+/** A user message that says something, not one that only returns tool results. */
+function isUserQuestion(message) {
+    if (!isUserMessage(message)) {
+        return false;
+    }
+    return (("content" in message && typeof message.content === "string") ||
+        messageBlocks(message).some((block) => blockField(block, "type") === "text"));
+}
+/**
+ * Removes every tool result whose call is not among `messages`, including one
+ * that shares a message with text, and drops messages left with nothing.
+ */
+function withoutUnansweredResults(messages) {
+    const calls = new Set(messages
+        .flatMap(messageBlocks)
+        .filter((block) => blockField(block, "type") === "tool_use")
+        .map((block) => blockField(block, "id")));
+    return messages.flatMap((message) => {
+        const blocks = messageBlocks(message);
+        const answered = blocks.filter((block) => blockField(block, "type") !== "tool_result" ||
+            calls.has(blockField(block, "tool_use_id")));
+        if (answered.length === blocks.length) {
+            return [message];
+        }
+        if (answered.length === 0 ||
+            message === null ||
+            typeof message !== "object") {
+            return [];
+        }
+        return [{ ...message, content: answered }];
+    });
+}
 /**
  * Serialize a model request for a span attribute without losing the
  * conversation to the cap.
  *
  * A head cut keeps the system prompt and drops everything after it, so a long
  * system prompt hides every message, the user's question included. Here the
- * prompt and the newest messages are budgeted first, older messages are
- * dropped whole and counted, and the system prompt gets whatever room is left.
- * The result is always valid JSON.
+ * prompt, the current question and this turn's newest messages are budgeted
+ * first, earlier messages are dropped whole and counted, and the system prompt
+ * gets whatever room is left. A tool result is never kept without the call it
+ * answers. The result is always valid JSON.
  */
-export function spanModelRequestAttribute(rawRequest, maxChars = SPAN_ATTRIBUTE_MAX_CHARS) {
+export function spanModelRequestAttribute(rawRequest, maxChars = spanAttributeMaxChars()) {
     // Credentials are masked before anything is measured, so budgets apply to
     // what is actually written. Each message is masked on its own: one that
     // cannot be walked is omitted without losing the rest of the conversation.
@@ -410,22 +481,83 @@ export function spanModelRequestAttribute(rawRequest, maxChars = SPAN_ATTRIBUTE_
         prompt = fitText(request.prompt, remaining);
         remaining -= jsonLength(prompt);
     }
+    // Shortens a copy of the message's longest strings; the shape survives.
+    const fitMessage = (message, budget) => {
+        if (jsonLength(message) <= budget) {
+            return message;
+        }
+        let copy;
+        try {
+            copy = JSON.parse(serialize(message));
+        }
+        catch {
+            copy = undefined;
+        }
+        if (copy !== null && typeof copy === "object") {
+            for (let attempt = 0; jsonLength(copy) > budget && attempt < MAX_SHRINK_ATTEMPTS; attempt++) {
+                if (!shrinkLongestString(copy, jsonLength(copy) - budget)) {
+                    break;
+                }
+            }
+            if (jsonLength(copy) <= budget) {
+                return copy;
+            }
+        }
+        return {
+            truncatedMessage: fitText(serialize(message), Math.max(0, budget - 32)),
+        };
+    };
     const messages = request.messages ?? [];
-    const kept = [];
+    // The current question is the newest user message with something to say;
+    // everything after it is this turn's tool calls and their results.
+    let questionIndex = -1;
     for (let index = messages.length - 1; index >= 0; index--) {
+        if (isUserQuestion(messages[index])) {
+            questionIndex = index;
+            break;
+        }
+    }
+    const questionFloor = questionIndex >= 0
+        ? Math.min(jsonLength(messages[questionIndex]) + 1, QUESTION_FLOOR_CHARS + TRUNCATION_NOTE_ALLOWANCE_CHARS)
+        : 0;
+    remaining -= questionFloor;
+    const turn = [];
+    let turnComplete = true;
+    for (let index = messages.length - 1; index > questionIndex; index--) {
         const size = jsonLength(messages[index]) + 1;
         if (size <= remaining) {
-            kept.unshift(messages[index]);
+            turn.unshift(messages[index]);
             remaining -= size;
             continue;
         }
-        if (kept.length === 0 && remaining > 64) {
+        turnComplete = false;
+        if (turn.length === 0 && remaining > 64) {
             const excerpt = fitText(serialize(messages[index]), remaining - 32);
-            kept.unshift({ truncatedMessage: excerpt });
+            turn.unshift({ truncatedMessage: excerpt });
             remaining -= jsonLength(excerpt) + 32;
         }
         break;
     }
+    remaining += questionFloor;
+    let kept = [];
+    if (questionIndex >= 0) {
+        const question = fitMessage(messages[questionIndex], remaining - 1);
+        remaining -= jsonLength(question) + 1;
+        const history = [];
+        for (let index = questionIndex - 1; turnComplete && index >= 0; index--) {
+            const size = jsonLength(messages[index]) + 1;
+            if (size > remaining) {
+                break;
+            }
+            history.unshift(messages[index]);
+            remaining -= size;
+        }
+        kept.push(...history, question);
+    }
+    kept.push(...turn);
+    if (kept.length < messages.length) {
+        kept = withoutUnansweredResults(kept);
+    }
     const omittedMessages = messages.length - kept.length;
     remaining += systemFloor;
     const body = {};

package/dist/telemetry/index.d.ts

@@ -1,7 +1,7 @@
 export { TelemetryService } from "./telemetryService.js";
 export { tracers } from "./tracers.js";
 export { withSpan, withClientSpan, withStreamSpan, withClientStreamSpan, } from "./withSpan.js";
-export { ATTR, LANGFUSE_ATTR, SPAN_ATTRIBUTE_MAX_CHARS, redactSpanContent, spanContentAttribute, spanJsonAttribute, spanModelRequestAttribute, } from "./attributes.js";
+export { ATTR, LANGFUSE_ATTR, SPAN_ATTRIBUTE_MAX_CHARS, redactSpanContent, spanAttributeMaxChars, spanContentAttribute, spanJsonAttribute, spanModelRequestAttribute, } from "./attributes.js";
 /**
  * Initialize telemetry for NeuroLink
  * Reuses an existing global TracerProvider when one is already registered,

package/dist/telemetry/index.js

@@ -2,7 +2,7 @@
 export { TelemetryService } from "./telemetryService.js";
 export { tracers } from "./tracers.js";
 export { withSpan, withClientSpan, withStreamSpan, withClientStreamSpan, } from "./withSpan.js";
-export { ATTR, LANGFUSE_ATTR, SPAN_ATTRIBUTE_MAX_CHARS, redactSpanContent, spanContentAttribute, spanJsonAttribute, spanModelRequestAttribute, } from "./attributes.js";
+export { ATTR, LANGFUSE_ATTR, SPAN_ATTRIBUTE_MAX_CHARS, redactSpanContent, spanAttributeMaxChars, spanContentAttribute, spanJsonAttribute, spanModelRequestAttribute, } from "./attributes.js";
 import { logger } from "../utils/logger.js";
 /**
  * Initialize telemetry for NeuroLink

package/dist/types/classifierRouter.d.ts

@@ -23,13 +23,17 @@ export type ClassifierDifficulty = "trivial" | "simple" | "moderate" | "hard" |
  *
  * - `heuristic` — keyword/length scoring. Deterministic, zero latency.
  * - `llm` — a cheap classifier model via the injected `generate`.
- * - `jev` — TypeSafe's System One model; one ~400ms round trip that returns a
- *   *calibrated* confidence rather than a self-reported one.
+ * - `jev` — a decision model, whichever decision provider is configured
+ *   (TypeSafe's Jev, Laya, XOR, SubQ, Perplexity, Cloudflare Clef or OpenAI
+ *   Decisions); one round trip that returns the confidence the provider
+ *   reports (*calibrated* on TypeSafe's Jev, ~400ms there) rather than a
+ *   self-reported one.
  * - `auto` — `jev` when a decision provider is configured, in the environment
  *   or in SDK credentials (`TYPESAFE_API_KEY`, `LAYA_API_KEY` with
- *   `LAYA_BASE_URL`, `XOR_API_KEY` with `XOR_BASE_URL`, or
- *   `PERPLEXITY_API_KEY`), otherwise
- *   `heuristic`.
+ *   `LAYA_BASE_URL`, `XOR_API_KEY` with `XOR_BASE_URL`, `SUBQ_API_KEY`, `PERPLEXITY_API_KEY`,
+ *   `CLOUDFLARE_API_KEY` with `CLOUDFLARE_ACCOUNT_ID`, or `OPENAI_API_KEY`,
+ *   which does not count while `OPENAI_BASE_URL`, `OPENAI_API_BASE` or `credentials.openai.baseURL` names a
+ *   host other than `https://api.openai.com`), otherwise `heuristic`.
  */
 export type ClassifierStrategyKind = "heuristic" | "llm" | "jev" | "auto";
 /**
@@ -190,10 +194,18 @@ export type ClassifierRouterConfig = {
      * Classification strategy. Default: "auto" — which resolves to "jev" when a
      * decision provider is configured, in the environment or in SDK credentials
      * (`TYPESAFE_API_KEY`, `LAYA_API_KEY` with `LAYA_BASE_URL`,
-     * `XOR_API_KEY` with `XOR_BASE_URL`, or `PERPLEXITY_API_KEY`) and "heuristic"
-     * otherwise, so configuring one upgrades routing without any code change.
-     * Behaviour for callers with no key is unchanged. `PERPLEXITY_API_KEY` is also
-     * the Perplexity text provider's key, so it counts here too.
+     * `XOR_API_KEY` with `XOR_BASE_URL`, `SUBQ_API_KEY`, `PERPLEXITY_API_KEY`,
+     * `CLOUDFLARE_API_KEY` with `CLOUDFLARE_ACCOUNT_ID`, or `OPENAI_API_KEY`) and
+     * "heuristic" otherwise, so configuring one upgrades routing without any code
+     * change. Behaviour for a caller with none of these is unchanged. Three of them
+     * are shared with a text provider and count here too: `PERPLEXITY_API_KEY`
+     * (Perplexity's), the Cloudflare pair (Workers AI's) and `OPENAI_API_KEY` (the
+     * OpenAI chat provider's). A caller whose only key is `OPENAI_API_KEY` now
+     * resolves to "jev", answered by OpenAI Decisions as the last-resort decision
+     * provider; each classification is then a paid call that still falls back to
+     * the heuristic on any failure. The exception is a key that sits beside an
+     * `OPENAI_BASE_URL`, `OPENAI_API_BASE` or `credentials.openai.baseURL` naming a host other than
+     * `https://api.openai.com`: that key is a gateway's and does not count.
      */
     classifier?: ClassifierStrategyKind;
     /** Model used by the "llm" strategy. Defaults to provider/model auto. */

package/dist/types/cli.d.ts

@@ -1780,8 +1780,10 @@ export type CliClassifierRouterFlags = {
      * Strategy: "auto" (default), "heuristic", "llm" or "jev"
      * (--classifier-strategy). "auto" resolves to "jev" when a decision
      * provider is configured (TYPESAFE_API_KEY, LAYA_API_KEY with
-     * LAYA_BASE_URL, XOR_API_KEY with XOR_BASE_URL, or PERPLEXITY_API_KEY) and
-     * "heuristic" otherwise.
+     * LAYA_BASE_URL, XOR_API_KEY with XOR_BASE_URL, SUBQ_API_KEY, PERPLEXITY_API_KEY,
+     * CLOUDFLARE_API_KEY with CLOUDFLARE_ACCOUNT_ID, or OPENAI_API_KEY, which does
+     * not count beside an OPENAI_BASE_URL, OPENAI_API_BASE or credentials.openai.baseURL naming a host other
+     * than https://api.openai.com) and "heuristic" otherwise.
      */
     classifierStrategy?: string;
     /**

package/dist/types/decision.d.ts

@@ -523,11 +523,70 @@ export type DecisionBooleanGate = {
     minConfidence?: number;
 };
 /** The wire family a decide provider speaks. */
-export type DecideDialectName = "system-one";
+export type DecideDialectName = "system-one" | "openai-decisions";
 export type DecisionAnswerReading = {
     readonly answers: Readonly<Record<string, DecisionAnswer>>;
     /** Ids of answers the server returned that could not be parsed. */
     readonly dropped: readonly string[];
+    /** Ids of questions the model declined to answer; they have no answer. */
+    readonly refused?: readonly string[];
+};
+/**
+ * The request body keys prepared media is written under. Set together with
+ * `decisionLimits.media`, which says how much of it a request may carry.
+ */
+export type DecideMedia = {
+    /** Written only when the request carries at least one image. */
+    imagesKey: string;
+    /** Written only when the request carries a video; required exactly when `decisionLimits.media.video` is true. */
+    videoKey?: string;
+};
+/**
+ * Says when a decide provider's key may be used as the default decision
+ * provider. A key found only in the environment may belong to something else:
+ * `OPENAI_API_KEY` is also the key of every OpenAI-compatible gateway, and
+ * sending a gateway's key and the decision state to the vendor's own host would
+ * disclose both. With a guard, the environment key is skipped for default
+ * resolution while any variable in `baseURLEnvVars`, or any credentials field in
+ * `credentialBaseURLs`, holds a URL whose origin (scheme, host and port,
+ * compared case-insensitively; path and trailing slash ignored) is not in
+ * `trustedOrigins`; a value that does not parse as an http(s) URL counts as
+ * untrusted, and a blank one as unset.
+ *
+ * It does not apply once the host routed the provider on purpose: the
+ * provider's own base-URL variable is set, or `credentials.<key>` supplies a key
+ * or a base URL. It never applies to a call that names the provider.
+ */
+export type DecideAmbientKeyGuard = {
+    /** Environment variables that move the key's other users off the vendor's host. */
+    baseURLEnvVars: readonly string[];
+    /** Origins that are the vendor's own, such as `https://api.openai.com`. */
+    trustedOrigins: readonly string[];
+    /**
+     * Credentials fields that move the key's other users off the vendor's host,
+     * read from the same `credentials` the resolver is given (the instance's
+     * merged with the call's), such as the chat provider's `credentials.openai`
+     * `baseURL`. Omitted means only the environment is read.
+     */
+    credentialBaseURLs?: readonly DecideAmbientKeyCredentialField[];
+};
+/** One credentials field a {@link DecideAmbientKeyGuard} reads: `credentials[key][field]`. */
+export type DecideAmbientKeyCredentialField = {
+    /** The `NeurolinkCredentials` key of the slice, such as `openai`. */
+    key: string;
+    /** The field in that slice. Only `baseURL` is read today. */
+    field: "baseURL";
+};
+/** What a dialect lays out as one request body. */
+export type DecisionDialectBody = {
+    readonly state: DecisionState;
+    /** Each question already written by the dialect's `encodeQuestion`, by id. */
+    readonly questions: Readonly<Record<string, Record<string, unknown>>>;
+    /** Undefined means the request names no model. */
+    readonly model?: string;
+    readonly media?: DecisionPreparedMedia;
+    /** Where a layout with named media fields puts them; other layouts ignore it. */
+    readonly mediaKeys?: Readonly<DecideMedia>;
 };
 /**
  * How questions are written to, and answers read from, a decide wire. One
@@ -536,6 +595,11 @@ export type DecisionAnswerReading = {
 export type DecisionDialect = {
     readonly name: DecideDialectName;
     encodeQuestion: (question: DecisionQuestion) => Record<string, unknown>;
-    /** `undefined` means the response has no usable answers container. */
-    readAnswers: (decoded: Readonly<Record<string, unknown>>, reportedConfidence: Readonly<Record<string, number>>) => DecisionAnswerReading | undefined;
+    buildBody: (input: DecisionDialectBody) => Record<string, unknown>;
+    /**
+     * `undefined` means the response has no usable answers container. `asked`
+     * is the request's own question map, for a wire that echoes names rather
+     * than the ids sent.
+     */
+    readAnswers: (decoded: Readonly<Record<string, unknown>>, reportedConfidence: Readonly<Record<string, number>>, asked: DecisionQuestionMap) => DecisionAnswerReading | undefined;
 };

package/dist/types/providerCatalog.d.ts

@@ -4,7 +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";
+import type { DecideAmbientKeyGuard, DecideDialectName, DecideMedia, DecisionErrorKind, DecisionLimits } from "./decision.js";
 export type CatalogPricingPerMTok = {
     input: number;
     output: number;
@@ -300,16 +300,6 @@ export type DecideModel = {
     /** With `body-unless-alias`: the model name that means "send no model". */
     alias?: string;
 };
-/**
- * The request body keys prepared media is written under. Set together with
- * `decisionLimits.media`, which says how much of it a request may carry.
- */
-export type DecideMedia = {
-    /** Written only when the request carries at least one image. */
-    imagesKey: string;
-    /** Written only when the request carries a video; required exactly when `decisionLimits.media.video` is true. */
-    videoKey?: string;
-};
 export type DecideBlock = {
     dialect: DecideDialectName;
     /** Decision order, lower runs first; unique across decide providers. */
@@ -325,6 +315,12 @@ export type DecideBlock = {
     model: DecideModel;
     /** Body keys for images and video, after `model`, `state` and `questions`. Omitted means the body never carries media. */
     media?: DecideMedia;
+    /**
+     * Keeps a key found only in the environment from becoming the default
+     * decision provider when it plainly belongs to something else. Omitted means
+     * no check.
+     */
+    ambientKeyGuard?: DecideAmbientKeyGuard;
     requestIdHeaders?: readonly string[];
     response?: {
         modelPaths?: readonly string[];

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" | "laya" | "lemonfox-ai" | "lilac" | "llmtech" | "longcat" | "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" | "runinfra" | "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" | "xiaomi-mimo" | "xor" | "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" | "longcat" | "mancer" | "metaModelApi" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshotAi" | "morph" | "naverClovaStudio" | "nebius" | "neuralwatt" | "novita" | "ovhcloud" | "parasail" | "paretoInference" | "perplexity" | "poolside" | "primeIntellect" | "reka" | "runinfra" | "sakanaAi" | "sambanova" | "sarvam" | "scaleway" | "siliconflow" | "stackit" | "stepfun" | "subconscious" | "synthetic" | "telnyx" | "thinkingMachines" | "tinfoil" | "together" | "umansAi" | "upstage" | "veniceAi" | "vispark" | "vultrInference" | "wafer" | "wandbInference" | "xai" | "xiaomiMimo" | "xor" | "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" | "longcat" | "mancer" | "meta-model-api" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshot-ai" | "morph" | "naver-clova-studio" | "nebius" | "neuralwatt" | "novita" | "openai-decisions" | "ovhcloud" | "parasail" | "pareto-inference" | "perplexity" | "poolside" | "prime-intellect" | "reka" | "runinfra" | "sakana-ai" | "sambanova" | "sarvam" | "scaleway" | "siliconflow" | "stackit" | "stepfun" | "subconscious" | "subq" | "synthetic" | "telnyx" | "thinking-machines" | "tinfoil" | "together-ai" | "umans-ai" | "upstage" | "venice-ai" | "vispark" | "vultr-inference" | "wafer" | "wandb-inference" | "xai" | "xiaomi-mimo" | "xor" | "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" | "longcat" | "mancer" | "metaModelApi" | "minimax" | "mistral" | "moark" | "modelscope" | "moonshotAi" | "morph" | "naverClovaStudio" | "nebius" | "neuralwatt" | "novita" | "openaiDecisions" | "ovhcloud" | "parasail" | "paretoInference" | "perplexity" | "poolside" | "primeIntellect" | "reka" | "runinfra" | "sakanaAi" | "sambanova" | "sarvam" | "scaleway" | "siliconflow" | "stackit" | "stepfun" | "subconscious" | "subq" | "synthetic" | "telnyx" | "thinkingMachines" | "tinfoil" | "together" | "umansAi" | "upstage" | "veniceAi" | "vispark" | "vultrInference" | "wafer" | "wandbInference" | "xai" | "xiaomiMimo" | "xor" | "zAi";

package/dist/types/providers.d.ts

@@ -6,7 +6,7 @@ import type { NeuroLink } from "../neurolink.js";
 import { AIProviderName, AnthropicModels, BedrockModels, DeepSeekModels, GoogleAIModels, LlamaCppModels, LMStudioModels, NvidiaNimModels, OpenAIModels, VertexModels } from "../constants/enums.js";
 import type { ValidationSchema } from "./aliases.js";
 import type { STTCredentials } from "./transcribe.js";
-import type { DecisionLimits, DecisionRequest, DecisionResult } from "./decision.js";
+import type { DecideAmbientKeyGuard, DecisionLimits, DecisionRequest, DecisionResult } from "./decision.js";
 import type { FileToolRootPolicy } from "./fileToolRoots.js";
 import type { EnhancedGenerateResult, TextGenerationOptions } from "./generate.js";
 import type { MultimodalAudioEntry, MultimodalVideoEntry } from "./file.js";
@@ -421,6 +421,10 @@ export type NeurolinkCredentials = {
         apiKey?: string;
         baseURL?: string;
     };
+    openaiDecisions?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     ovhcloud?: {
         apiKey?: string;
         baseURL?: string;
@@ -485,6 +489,10 @@ export type NeurolinkCredentials = {
         apiKey?: string;
         baseURL?: string;
     };
+    subq?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     synthetic?: {
         apiKey?: string;
         baseURL?: string;
@@ -2461,6 +2469,13 @@ export type ProviderDescriptor = {
      * descriptor that serves `decide`; checked when the descriptors load.
      */
     decideOrder?: number;
+    /**
+     * Skips this provider's environment key for default resolution while a
+     * listed base-URL variable names a host outside the trusted origins. Only a
+     * decide provider whose key is also another provider's can use it. See
+     * {@link DecideAmbientKeyGuard}.
+     */
+    ambientKeyGuard?: DecideAmbientKeyGuard;
     /**
      * Ascending priority (1 = tried first) in the auto-select fallback chain
      * used by getBestProvider(). Undefined = not part of the auto-select chain.

package/dist/utils/pricing.js

@@ -803,6 +803,15 @@ const PRICING = {
         _default: { input: 0.04 / 1_000_000, output: 0 },
         "pplx-decider-v1-27b": { input: 0.04 / 1_000_000, output: 0 },
     },
+    "openai-decisions": {
+        // $0.10 per million input tokens, input only: `output: 0` is a price, not a
+        // token count, so do not "correct" it from `usage.output_tokens`. Source:
+        // https://developers.openai.com/api/docs/guides/decisions (read 2026-10-07).
+        // Regional and long-context multipliers are not modelled. Only `_default`:
+        // LiteLLM, OpenRouter and openai-compatible search every table by model
+        // name, so a `"gpt-6-luna"` key would price that chat model at this rate.
+        _default: { input: 0.1 / 1_000_000, output: 0 },
+    },
     "cloudflare-clef": {
         // Workers AI bills Clef per million INPUT tokens ($0.24 for clef, $0.09 for
         // clef-flash); no output price is listed. `output: 0` is a price, not a token