@juspay/neurolink 12.46.1 → 12.47.1

package/dist/cli/commands/decide.js

@@ -191,7 +191,7 @@ 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 and Perplexity take up to 8)",
+        describe: "Image for the model to read: a file path or a data: URL. Repeat for several (XOR and Perplexity take up to 8, Cloudflare Clef up to 4)",
     })
         .option("video", {
         type: "string",
@@ -217,7 +217,7 @@ 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 --image ./photo.png --questions \'{"color":{"type":"choice","instructions":"What color is the image?","criteria":{"red":"red","blue":"blue"}}}\'', "Ask about an image (XOR and Perplexity read images)"),
+        .example('$0 decide "What color is this?" --provider xor --image ./photo.png --questions \'{"color":{"type":"choice","instructions":"What color is the image?","criteria":{"red":"red","blue":"blue"}}}\'', "Ask about an image (XOR, Perplexity and Cloudflare Clef read images)"),
     handler: async (argv) => {
         const outputFormat = argv.format ?? "text";
         // --- Validate before any provider work ---------------------------------

package/dist/cli/commands/setup.js

@@ -19,8 +19,9 @@ import { handleBedrockSetup } from "./setup-bedrock.js";
 import { handleGCPSetup } from "./setup-gcp.js";
 import { handleHuggingFaceSetup } from "./setup-huggingface.js";
 import { handleMistralSetup } from "./setup-mistral.js";
+import { OpenRouterModels, } from "../../constants/enums.js";
 import { PROVIDER_DESCRIPTORS_BY_NAME } from "../../factories/providerDescriptors.js";
-import { createCohereConfig, createDeepSeekConfig, createIdeogramConfig, createJinaConfig, createNvidiaNimConfig, createOpenAICompatibleConfig, createRecraftConfig, createReplicateConfig, createStabilityConfig, createVoyageConfig, createTypeSafeConfig, createLayaConfig, createXorConfig, createPerplexityDeciderConfig, satisfiesFallbacks, } from "../../utils/providerConfig.js";
+import { createCohereConfig, createDeepSeekConfig, createIdeogramConfig, createJinaConfig, createNvidiaNimConfig, createOpenAICompatibleConfig, createRecraftConfig, createReplicateConfig, createStabilityConfig, createVoyageConfig, createTypeSafeConfig, createLayaConfig, createXorConfig, createPerplexityDeciderConfig, createCloudflareClefConfig, satisfiesFallbacks, } from "../../utils/providerConfig.js";
 import { getCatalogJsonEntries, buildCatalogConfigOptions, } from "../../providers/catalog/loader.js";
 // Provider information database
 const PROVIDERS = [
@@ -172,6 +173,7 @@ export const EXTRA_PROVIDER_CONFIGS = {
     laya: createLayaConfig(),
     xor: createXorConfig(),
     "perplexity-decider": createPerplexityDeciderConfig(),
+    "cloudflare-clef": createCloudflareClefConfig(),
     ...Object.fromEntries(getCatalogJsonEntries()
         .filter((e) => e.id !== "mistral")
         .map((e) => [e.id, buildCatalogConfigOptions(e)])),
@@ -570,13 +572,13 @@ async function handleOpenRouterSetup() {
     logger.always(chalk.cyan("  export OPENROUTER_API_KEY=your_api_key_here"));
     logger.always("");
     logger.always(chalk.yellow("Step 3: Test the configuration"));
-    logger.always(chalk.cyan('  neurolink generate "Hello!" --provider openrouter --model google/gemini-2.0-flash-exp:free'));
+    logger.always(chalk.cyan(`  neurolink generate "Hello!" --provider openrouter --model ${OpenRouterModels.GEMINI_2_5_FLASH}`));
     logger.always("");
     logger.always(chalk.green("Available models include:"));
-    logger.always("  • anthropic/claude-3.5-sonnet - Best for analysis");
-    logger.always("  • openai/gpt-4o - Industry standard");
-    logger.always("  • google/gemini-2.0-flash-exp:free - Free tier");
-    logger.always("  • meta-llama/llama-3.1-70b-instruct - Open source");
+    logger.always(`  • ${OpenRouterModels.CLAUDE_SONNET_4_6} - Best for analysis`);
+    logger.always(`  • ${OpenRouterModels.GPT_4O} - Industry standard`);
+    logger.always(`  • ${OpenRouterModels.GEMINI_2_5_FLASH} - Fast and efficient`);
+    logger.always(`  • ${OpenRouterModels.LLAMA_3_1_70B} - Open source`);
     logger.always("");
     logger.always(chalk.gray("See all models at: https://openrouter.ai/models"));
 }

package/dist/cli/factories/commandFactory.js

@@ -616,7 +616,7 @@ export class CLICommandFactory {
         },
         classifierStrategy: {
             type: "string",
-            description: "Classifier strategy: 'auto' (default — 'jev' when a decision provider is configured, such as TYPESAFE_API_KEY, LAYA_API_KEY with LAYA_BASE_URL, XOR_API_KEY with XOR_BASE_URL, or PERPLEXITY_API_KEY, else 'heuristic'), 'heuristic' (no LLM), 'llm' (a cheap model picks per prompt), or 'jev' (a System One decision model — TypeSafe Jev, Laya, XOR or Perplexity — with calibrated confidence).",
+            description: "Classifier strategy: 'auto' (default — 'jev' when a decision provider is configured, such as TYPESAFE_API_KEY, LAYA_API_KEY with LAYA_BASE_URL, XOR_API_KEY with XOR_BASE_URL, PERPLEXITY_API_KEY, or CLOUDFLARE_API_KEY with CLOUDFLARE_ACCOUNT_ID, else 'heuristic'), 'heuristic' (no LLM), 'llm' (a cheap model picks per prompt), or 'jev' (a System One decision model — TypeSafe Jev, Laya, XOR, Perplexity or Cloudflare Clef — with calibrated confidence).",
             choices: ["auto", "heuristic", "llm", "jev"],
             alias: "classifier-strategy",
         },

package/dist/cli/proxy-clients/copilot.js

@@ -46,12 +46,66 @@ function getCopilotEnvPath() {
 function getProfileCandidates() {
     return [".zshrc", ".zprofile", ".bashrc", ".bash_profile", ".profile"].map((name) => join(homedir(), name));
 }
+/** A command that is `source`/`.` of a path ending in exactly `copilot-env.sh`, optionally after `then`/`else`/`do`. */
+const ENV_SCRIPT_SOURCE_COMMAND = /^\s*(?:(?:then|else|do)\s+)?(?:source|\.)\s+(["']?)(?:[^\s"']*\/)?copilot-env\.sh\1(?:\s|$)/;
+/**
+ * One profile line as the commands it runs, comment dropped. Quote-aware, so a
+ * `;` or `#` inside quotes neither splits nor starts a comment. Heredocs,
+ * multi-line strings and line continuations are not modelled, so a source line
+ * inside one is still counted, and a source built with `eval` is missed. Both
+ * are rare; this guards against the common mentions (a comment, an `echo`, a
+ * file-existence test, a `.bak` name) that suppressed the note before.
+ */
+function splitProfileCommands(line) {
+    const commands = [];
+    let current = "";
+    let quote = null;
+    for (let i = 0; i < line.length; i++) {
+        const char = line[i];
+        if (quote) {
+            current += char;
+            if (quote === '"' && char === "\\") {
+                current += line[i + 1] ?? "";
+                i++;
+            }
+            else if (char === quote) {
+                quote = null;
+            }
+            continue;
+        }
+        if (char === "\\") {
+            current += char + (line[i + 1] ?? "");
+            i++;
+            continue;
+        }
+        if (char === '"' || char === "'") {
+            quote = char;
+            current += char;
+            continue;
+        }
+        if (char === "#" &&
+            (current === "" || /\s/.test(current[current.length - 1]))) {
+            break;
+        }
+        if (char === ";" || char === "&" || char === "|") {
+            commands.push(current);
+            current = "";
+            continue;
+        }
+        current += char;
+    }
+    commands.push(current);
+    return commands;
+}
 /** Whether any shell profile already sources the generated script. */
 async function isEnvScriptSourced() {
     const fs = await import("fs");
     for (const profile of getProfileCandidates()) {
         try {
-            if (fs.readFileSync(profile, "utf8").includes("copilot-env.sh")) {
+            if (fs
+                .readFileSync(profile, "utf8")
+                .split(/\r?\n/)
+                .some((line) => splitProfileCommands(line).some((command) => ENV_SCRIPT_SOURCE_COMMAND.test(command)))) {
                 return true;
             }
         }

package/dist/constants/contextWindows.js

@@ -122,6 +122,11 @@ export const MODEL_CONTEXT_WINDOWS = {
         "claude-fable-5-1": 1_000_000,
         // Claude 5 (mid 2026) — 1M context window
         "claude-sonnet-5": 1_000_000,
+        "claude-opus-5": 1_000_000,
+        "claude-fable-5": 1_000_000,
+        // Claude 4.7 / 4.8 — 1M context window (platform.claude.com/docs/en/models)
+        "claude-opus-4-8": 1_000_000,
+        "claude-opus-4-7": 1_000_000,
         // Claude 4.6 (Feb 2026) — 1M context window
         "claude-opus-4-6": 1_000_000,
         "claude-sonnet-4-6": 1_000_000,
@@ -220,6 +225,13 @@ export const MODEL_CONTEXT_WINDOWS = {
         "claude-sonnet-5-5": 1_000_000,
         "claude-fable-5-1": 1_000_000,
         "claude-sonnet-5": 1_000_000,
+        // Placed before the 200K `claude-opus-4` prefix key below: the lookup takes
+        // the first prefix that matches, so a dated `claude-opus-4-8@...` id would
+        // otherwise land on 200K.
+        "claude-opus-5": 1_000_000,
+        "claude-fable-5": 1_000_000,
+        "claude-opus-4-8": 1_000_000,
+        "claude-opus-4-7": 1_000_000,
         "claude-opus-4-6": 1_000_000,
         "claude-sonnet-4-6": 1_000_000,
         "claude-sonnet-4-5": 200_000,

package/dist/constants/enums.d.ts

@@ -115,6 +115,12 @@ export declare enum AIProviderName {
      * inference type only. Distinct from `PERPLEXITY`, the Sonar text provider.
      */
     PERPLEXITY_DECIDER = "perplexity-decider",
+    /**
+     * Cloudflare Clef (`@cf/cloudflare/clef`, `@cf/cloudflare/clef-flash`) on
+     * Workers AI — serves the `decide` inference type only. Distinct from
+     * `CLOUDFLARE`, the Workers AI text provider.
+     */
+    CLOUDFLARE_CLEF = "cloudflare-clef",
     AUTO = "auto"
 }
 /**
@@ -1781,3 +1787,13 @@ export declare enum XorModels {
 export declare enum PerplexityDeciderModels {
     PPLX_DECIDER_V1_27B = "pplx-decider-v1-27b"
 }
+/**
+ * Cloudflare Clef decision models, named as the Workers AI API names them in
+ * the request body (the path adds the `@cf/cloudflare/` prefix). Hand-written:
+ * Clef is a Tier-3 provider, so it is not in the provider catalog and codegen
+ * never touches this. `CLEF` is the 27B model, `CLEF_FLASH` the 9B one.
+ */
+export declare enum CloudflareClefModels {
+    CLEF = "clef",
+    CLEF_FLASH = "clef-flash"
+}

package/dist/constants/enums.js

@@ -121,6 +121,12 @@ export var AIProviderName;
      * inference type only. Distinct from `PERPLEXITY`, the Sonar text provider.
      */
     AIProviderName["PERPLEXITY_DECIDER"] = "perplexity-decider";
+    /**
+     * Cloudflare Clef (`@cf/cloudflare/clef`, `@cf/cloudflare/clef-flash`) on
+     * Workers AI — serves the `decide` inference type only. Distinct from
+     * `CLOUDFLARE`, the Workers AI text provider.
+     */
+    AIProviderName["CLOUDFLARE_CLEF"] = "cloudflare-clef";
     AIProviderName["AUTO"] = "auto";
 })(AIProviderName || (AIProviderName = {}));
 /**
@@ -2103,3 +2109,14 @@ export var PerplexityDeciderModels;
 (function (PerplexityDeciderModels) {
     PerplexityDeciderModels["PPLX_DECIDER_V1_27B"] = "pplx-decider-v1-27b";
 })(PerplexityDeciderModels || (PerplexityDeciderModels = {}));
+/**
+ * Cloudflare Clef decision models, named as the Workers AI API names them in
+ * the request body (the path adds the `@cf/cloudflare/` prefix). Hand-written:
+ * Clef is a Tier-3 provider, so it is not in the provider catalog and codegen
+ * never touches this. `CLEF` is the 27B model, `CLEF_FLASH` the 9B one.
+ */
+export var CloudflareClefModels;
+(function (CloudflareClefModels) {
+    CloudflareClefModels["CLEF"] = "clef";
+    CloudflareClefModels["CLEF_FLASH"] = "clef-flash";
+})(CloudflareClefModels || (CloudflareClefModels = {}));

package/dist/core/toolExecutionGuards.d.ts

@@ -13,15 +13,7 @@
  */
 import type { Tool, ToolExecutionGuards } from "../types/index.js";
 /**
- * Mid-turn tool sync for the native Gemini loops that build their snapshot
- * via buildNativeToolDeclarations. `search_tools` (tools.discovery) hydrates
- * discovered tools into the live record between steps; without this refresh
- * they stay invisible to the rest of the turn and every call dies as
- * TOOL_NOT_FOUND. Mutates the snapshot in place — the request config holds
- * `toolsConfig` by reference — and returns true when anything was added.
- */
-/**
- * Everything a native Gemini loop wraps around a tool call that the shared
+ * Everything a native provider loop wraps around a tool call that the shared
  * engine does not do itself.
  *
  * Order matters. `raceWithAbort` sits INSIDE `withTimeout` so a turn-level

package/dist/core/toolExecutionGuards.js

@@ -13,15 +13,7 @@
  */
 import { raceWithAbort, withTimeout } from "../utils/async/index.js";
 /**
- * Mid-turn tool sync for the native Gemini loops that build their snapshot
- * via buildNativeToolDeclarations. `search_tools` (tools.discovery) hydrates
- * discovered tools into the live record between steps; without this refresh
- * they stay invisible to the rest of the turn and every call dies as
- * TOOL_NOT_FOUND. Mutates the snapshot in place — the request config holds
- * `toolsConfig` by reference — and returns true when anything was added.
- */
-/**
- * Everything a native Gemini loop wraps around a tool call that the shared
+ * Everything a native provider loop wraps around a tool call that the shared
  * engine does not do itself.
  *
  * Order matters. `raceWithAbort` sits INSIDE `withTimeout` so a turn-level

package/dist/factories/providerDescriptors.js

@@ -1,5 +1,5 @@
 import { AIProviderName } from "../constants/enums.js";
-import { GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, TypeSafeModels, LayaModels, XorModels, PerplexityDeciderModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, ReplicateModels, } from "../constants/enums.js";
+import { GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, TypeSafeModels, LayaModels, XorModels, PerplexityDeciderModels, CloudflareClefModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, ReplicateModels, } from "../constants/enums.js";
 import { API_KEY_FORMATS } from "../utils/providerConfig.js";
 import { getCatalogJsonEntries, catalogCredentialsKey, catalogEnvVar, } from "../providers/catalog/loader.js";
 import { DEFAULT_INFERENCE_KINDS } from "../types/index.js";
@@ -451,7 +451,8 @@ const HAND_DESCRIPTORS = [
         timeouts: { decideMs: 5000 },
         setupUrl: "https://console.typesafe.ai/keys",
     },
-    // Laya MUST stay after TypeSafe, XOR after Laya, and Perplexity after XOR.
+    // Laya MUST stay after TypeSafe, XOR after Laya, Perplexity after XOR, and
+    // Cloudflare Clef after Perplexity.
     // resolveDefaultDecisionProvider() returns the first configured
     // DECISION_PROVIDERS entry, in this order, so a host configured for several
     // keeps Jev for every built-in consumer and reaches the others only by
@@ -606,6 +607,79 @@ const HAND_DESCRIPTORS = [
         },
         setupUrl: "https://console.perplexity.ai",
     },
+    {
+        name: AIProviderName.CLOUDFLARE_CLEF,
+        aliases: [],
+        credentialsKey: "cloudflareClef",
+        envVars: {
+            // The same token and account id the `cloudflare` text provider reads, so
+            // ambient Workers AI settings configure this provider too. It sits last
+            // in this list, which is what keeps that from displacing any other
+            // decision provider a host has configured.
+            apiKey: "CLOUDFLARE_API_KEY",
+            // Cloudflare's own API is the endpoint, so a base URL is optional; but
+            // the path carries the account id, so a token alone is useless. Both are
+            // required for it to count as configured, from the environment or from
+            // credentials.cloudflareClef.
+            extraRequired: ["CLOUDFLARE_ACCOUNT_ID"],
+            extraRequiredCredentialFields: { CLOUDFLARE_ACCOUNT_ID: "accountId" },
+            baseURL: "CLOUDFLARE_CLEF_BASE_URL",
+            model: "CLOUDFLARE_CLEF_MODEL",
+        },
+        defaultModel: CloudflareClefModels.CLEF,
+        // Serves only `decide`, like the other decision providers — the one
+        // declaration that keeps it out of every generation code path, and out of
+        // the way of the `cloudflare` text provider.
+        inferenceKinds: ["decide"],
+        toolSupport: "none",
+        localRuntime: false,
+        healthCheck: "env-only",
+        // Deliberately NO autoSelectPriority / autoSelectPreference /
+        // defaultHealthSweepPriority, for the same reason as TypeSafe above.
+        //
+        // Measured 2026-10-03: 0.3 to 1.0 s for a small request, 1.1 s for 64 questions on
+        // clef-flash and 1.3 s on clef, and 2.2 s at most for any of 60 requests
+        // sent at once. On 2026-10-04, 64 questions took 1.5 s (flash) / 2.3 s (clef).
+        // 5s covers those observations and keeps a fail-open consumer
+        // from waiting on a stuck call.
+        timeouts: { decideMs: 5_000 },
+        // The Workers AI endpoint ignores state text past about 2,048 tokens
+        // (hosted service or model: unknown), despite the documented 64K. The local
+        // 1,500-token estimate refuses before every measured cut. On 2026-10-04,
+        // both models read facts at the original clef-flash lower bounds for logs,
+        // number lists, digit arrays and compact JSON, but not about 2.5% further on;
+        // English prose and random CJK already matched on both. Digits cost 1,
+        // ASCII symbols 0.75, BMP non-ASCII 1.5 and astral characters 3 tokens.
+        // Natural Chinese, Japanese, Korean and Hindi prose and emoji-rich English
+        // were also safe under those rates on clef. A many-key object cut between
+        // 128 and 134 preceding keys on both models: its lower bound was 4,883
+        // compact-JSON characters, estimated at 2,299 tokens. The probe used an
+        // explicit field-name question and zero-padded keys together after the
+        // control failed three times. It then passed; necessity was not established.
+        // Other object shapes and Unicode sequences may differ.
+        decisionLimits: {
+            maxStateTokens: 1_500,
+            maxQuestions: 64,
+            nonAsciiTokensPerChar: 1.5,
+            digitTokensPerChar: 1,
+            symbolTokensPerChar: 0.75,
+            astralTokensPerChar: 3,
+            // Images only. Exactly four tiny images succeeded and five were refused
+            // on both models. image/jpg also succeeded and is normalized to JPEG.
+            // Retain the conservative 256,000-byte encoded-body cap: on 2026-10-04
+            // both models accepted 520,000 text characters and refused 525,000 with
+            // 413/code 5021. On 2026-10-03, clef-flash accepted 262,000 and refused
+            // 270,000. Estimates match encoded body characters / 4, rounded up. The
+            // threshold moved from 65,527 accepted / 67,527 refused to 130,026 /
+            // 131,276 (clef) and 130,027 / 131,277 (flash); errors still print 65,536.
+            // Inference: the new interval contains 131,072 (twice the printed
+            // figure); reason unknown.
+            // The new image-byte ceiling was not measured; the old 195/202 KB PNG boundary
+            // is historical. See the guide for exact model and date coverage.
+            media: { maxImages: 4, video: false, maxRequestBytes: 256_000 },
+        },
+        setupUrl: "https://dash.cloudflare.com/profile/api-tokens",
+    },
 ];
 /**
  * Builds a ProviderDescriptor for every JSON-catalog provider. Every field
@@ -724,9 +798,14 @@ function isDecisionProviderConfigured(descriptor, credentials) {
     const hasKey = [descriptor.envVars.apiKey, ...(descriptor.envVars.fallbacks ?? [])].some(inEnv) ||
         isSet(slice?.apiKey) ||
         isSet(slice?.gatewayApiKey);
-    // A required base URL can also come from credentials.<key>.baseURL.
-    const hasRequired = (descriptor.envVars.extraRequired ?? []).every((name) => inEnv(name) ||
-        (name === descriptor.envVars.baseURL && isSet(slice?.baseURL)));
+    // A required base URL can also come from credentials.<key>.baseURL, and any
+    // other required value the descriptor maps to a field of that slice.
+    const hasRequired = (descriptor.envVars.extraRequired ?? []).every((name) => {
+        const field = descriptor.envVars.extraRequiredCredentialFields?.[name];
+        return (inEnv(name) ||
+            (name === descriptor.envVars.baseURL && isSet(slice?.baseURL)) ||
+            (field !== undefined && isSet(slice?.[field])));
+    });
     return hasKey && hasRequired;
 }
 /**

package/dist/factories/providerFactory.js

@@ -22,6 +22,17 @@ export function resolveCredentialKey(providerName) {
     const normalized = providerName.toLowerCase();
     return (ProviderFactory.getDescriptor(normalized)?.credentialsKey ?? normalized);
 }
+/**
+ * registerProvider() takes a class as well as a factory (ProviderConstructor),
+ * and calling a class without `new` throws. A prototype test cannot tell them
+ * apart, because every non-arrow function's prototype points back at itself,
+ * so the source text decides. A factory is therefore never constructed and
+ * never retried after it throws.
+ */
+function isClassConstructor(candidate) {
+    return (typeof candidate === "function" &&
+        /^class[\s{]/.test(Function.prototype.toString.call(candidate)));
+}
 /**
  * True Factory Pattern implementation for AI Providers
  * Uses registration-based approach to eliminate switch statements
@@ -113,7 +124,11 @@ export class ProviderFactory {
             if (typeof registration.constructor !== "function") {
                 throw new Error(`Invalid constructor for provider ${providerName}: not a function`);
             }
-            const factoryResult = registration.constructor(model, resolvedProviderName, sdk, region, scopedCredentials);
+            // A factory keeps being called as a method of the registration, as it
+            // always was; only a class is constructed.
+            const factoryResult = isClassConstructor(registration.constructor)
+                ? new registration.constructor(model, resolvedProviderName, sdk, region, scopedCredentials)
+                : registration.constructor(model, resolvedProviderName, sdk, region, scopedCredentials);
             const result = factoryResult instanceof Promise ? await factoryResult : factoryResult;
             return result;
         }

package/dist/factories/providerRegistry.js

@@ -1,6 +1,6 @@
 import { ProviderFactory } from "./providerFactory.js";
 import { logger } from "../utils/logger.js";
-import { AIProviderName, GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, TypeSafeModels, LayaModels, XorModels, PerplexityDeciderModels, ReplicateModels, } from "../constants/enums.js";
+import { AIProviderName, GoogleAIModels, OpenAIModels, AnthropicModels, VertexModels, OllamaModels, LiteLLMModels, NvidiaNimModels, OpenRouterModels, CohereModels, VoyageModels, JinaModels, StabilityModels, IdeogramModels, RecraftModels, TypeSafeModels, LayaModels, XorModels, PerplexityDeciderModels, CloudflareClefModels, ReplicateModels, } from "../constants/enums.js";
 import { PROVIDER_DESCRIPTORS_BY_NAME } from "./providerDescriptors.js";
 import { OPENAI_COMPAT_CATALOG } from "../providers/openaiCompatCatalog.js";
 import { providerChoicesFor } from "./mediaHandlerCatalog.js";
@@ -260,6 +260,15 @@ export class ProviderRegistry {
                 return new PerplexityDeciderProvider(modelName, sdk, undefined, perplexityDeciderCreds);
             }, process.env.PERPLEXITY_DECIDER_MODEL ||
                 PerplexityDeciderModels.PPLX_DECIDER_V1_27B, [], PROVIDER_DESCRIPTORS_BY_NAME.get(AIProviderName.PERPLEXITY_DECIDER));
+            // Register Cloudflare Clef — a `decide` provider reached through the
+            // Workers AI REST API. Its descriptor declares inferenceKinds:
+            // ["decide"], so nothing in the generation fallback chain can reach it,
+            // and it is registered apart from the `cloudflare` text provider.
+            ProviderFactory.registerProvider(AIProviderName.CLOUDFLARE_CLEF, async (modelName, _providerName, sdk, _region, credentials) => {
+                const cloudflareClefCreds = credentials;
+                const { CloudflareClefProvider } = await import("../providers/cloudflareClef.js");
+                return new CloudflareClefProvider(modelName, sdk, undefined, cloudflareClefCreds);
+            }, process.env.CLOUDFLARE_CLEF_MODEL || CloudflareClefModels.CLEF, [], PROVIDER_DESCRIPTORS_BY_NAME.get(AIProviderName.CLOUDFLARE_CLEF));
             logger.debug("All AI providers registered successfully");
             // ===== MEDIA HANDLER REGISTRATION =====
             // Single registration path (Task 11): each ecosystem barrel (voice,

package/dist/files/fileReferenceRegistry.js

@@ -827,7 +827,7 @@ export class FileReferenceRegistry {
                     extractedText = await this.extractWordText(buffer, ref);
                     break;
                 case "pptx":
-                    extractedText = await this.extractPptxText(buffer);
+                    extractedText = await this.extractPptxText(buffer, ref.filename);
                     break;
                 case "video":
                     extractedText = await this.extractVideoContent(buffer, ref);
@@ -992,13 +992,13 @@ export class FileReferenceRegistry {
     /**
      * Extract text from a PowerPoint file using PptxProcessor.
      */
-    async extractPptxText(buffer) {
+    async extractPptxText(buffer, filename) {
         try {
             const { PptxProcessor } = await import("../processors/document/PptxProcessor.js");
             return await PptxProcessor.extractText(buffer);
         }
         catch (err) {
-            logger.warn(`[FileReferenceRegistry] PPTX extraction failed: ${err instanceof Error ? err.message : String(err)}`);
+            logger.warn(`[FileReferenceRegistry] PPTX extraction failed for "${filename}": ${err instanceof Error ? err.message : String(err)}`);
             return null;
         }
     }

package/dist/models/manifestRegistry.js

@@ -23,6 +23,7 @@ import { typesafeManifest } from "./manifests/typesafe.js";
 import { layaManifest } from "./manifests/laya.js";
 import { xorManifest } from "./manifests/xor.js";
 import { perplexityDeciderManifest } from "./manifests/perplexityDecider.js";
+import { cloudflareClefManifest } from "./manifests/cloudflareClef.js";
 import { cohereManifest } from "./manifests/cohere.js";
 import { togetherAiManifest } from "./manifests/together-ai.js";
 import { fireworksManifest } from "./manifests/fireworks.js";
@@ -124,6 +125,7 @@ export const MANIFEST_REGISTRY = {
     laya: layaManifest,
     xor: xorManifest,
     "perplexity-decider": perplexityDeciderManifest,
+    "cloudflare-clef": cloudflareClefManifest,
     cerebras: catalogManifest("cerebras"),
     sambanova: catalogManifest("sambanova"),
     cohere: cohereManifest,

package/dist/models/manifests/cloudflareClef.d.ts

@@ -0,0 +1,16 @@
+import type { ProviderModelManifest } from "../../types/index.js";
+/**
+ * Cloudflare Clef on Workers AI — the `decide` inference type, not text
+ * generation, and not the Workers AI text models of the `cloudflare` provider.
+ *
+ * `contextWindow` is the figure Cloudflare documents (65,536 tokens). The
+ * Workers AI endpoint ignores state text past about 2,048 tokens without an
+ * error (hosted service or model: unknown), so NeuroLink enforces the limit on
+ * the descriptor's `decisionLimits`, not this. `vision` describes `generate()`
+ * input, which this provider does not serve; its image input is
+ * `DecisionRequest.images`.
+ *
+ * `maxOutputTokens` is required by the manifest type but has no honest value
+ * for a model that emits no text, so a small non-zero figure is used.
+ */
+export declare const cloudflareClefManifest: ProviderModelManifest;

package/dist/models/manifests/cloudflareClef.js

@@ -0,0 +1,42 @@
+/**
+ * Cloudflare Clef on Workers AI — the `decide` inference type, not text
+ * generation, and not the Workers AI text models of the `cloudflare` provider.
+ *
+ * `contextWindow` is the figure Cloudflare documents (65,536 tokens). The
+ * Workers AI endpoint ignores state text past about 2,048 tokens without an
+ * error (hosted service or model: unknown), so NeuroLink enforces the limit on
+ * the descriptor's `decisionLimits`, not this. `vision` describes `generate()`
+ * input, which this provider does not serve; its image input is
+ * `DecisionRequest.images`.
+ *
+ * `maxOutputTokens` is required by the manifest type but has no honest value
+ * for a model that emits no text, so a small non-zero figure is used.
+ */
+export const cloudflareClefManifest = {
+    defaultContextWindow: 65_536,
+    models: {
+        _default: {
+            aliases: [],
+            contextWindow: 65_536,
+            maxOutputTokens: 256,
+            vision: false,
+            functionCalling: false,
+        },
+        clef: {
+            aliases: ["@cf/cloudflare/clef"],
+            displayName: "Cloudflare Clef (27B)",
+            contextWindow: 65_536,
+            maxOutputTokens: 256,
+            vision: false,
+            functionCalling: false,
+        },
+        "clef-flash": {
+            aliases: ["@cf/cloudflare/clef-flash"],
+            displayName: "Cloudflare Clef-flash (9B)",
+            contextWindow: 65_536,
+            maxOutputTokens: 256,
+            vision: false,
+            functionCalling: false,
+        },
+    },
+};

package/dist/neurolink.js

@@ -8640,11 +8640,22 @@ Current user's request: ${currentInput}`;
                 }
                 yield fallbackChunk;
             }
+            // Read after the drain: a background-loop provider (native Vertex+Claude)
+            // fills these while the stream is consumed, so the values taken before
+            // the loop are still empty for it. Without this, a fallback that ran
+            // tools reports none on the result and in the end-of-turn events, and
+            // the 0-output gate below counts tool calls that had not appeared yet.
+            const finalToolCalls = fallbackResult.toolCalls ?? [];
+            const finalToolResults = fallbackResult.toolResults ?? [];
             if (fallbackRealOutputChunks === 0 &&
-                fallbackToolCalls.length === 0 &&
-                fallbackToolResults.length === 0) {
+                finalToolCalls.length === 0 &&
+                finalToolResults.length === 0) {
                 throw new Error(`Fallback provider ${fallbackRoute.provider} also returned 0 real output chunks (chunkCount=${fallbackChunkCount}, sentinel-only or empty)`);
             }
+            streamState.toolCalls = finalToolCalls;
+            streamState.toolResults = finalToolResults;
+            streamState.finishReason =
+                fallbackResult.finishReason ?? streamState.finishReason;
             // Fallback succeeded - likely guardrails blocked primary
             metadata.fallbackProvider = fallbackRoute.provider;
             metadata.fallbackModel = fallbackRoute.model;

package/dist/processors/media/VideoProcessor.js

@@ -660,6 +660,7 @@ export class VideoProcessor extends BaseFileProcessor {
      */
     async probeVideoWithFfmpeg(filePath) {
         let report;
+        let exitFailure;
         try {
             const { stderr } = await runFfmpeg(["-hide_banner", "-i", filePath, "-t", "0", "-f", "null", "-"], { timeoutMs: VIDEO_CONFIG.FFPROBE_TIMEOUT_MS });
             report = stderr;
@@ -668,18 +669,29 @@ export class VideoProcessor extends BaseFileProcessor {
             // A non-zero exit still carries the stream report when the failure came
             // after the input was opened.
             const stderr = error.stderr;
+            const message = error instanceof Error ? error.message : String(error);
             if (typeof stderr !== "string") {
                 return {
                     success: false,
-                    error: `ffmpeg probe failed: ${error instanceof Error ? error.message : String(error)}`,
+                    error: `ffmpeg probe failed: ${message}`,
                 };
             }
             report = stderr;
+            exitFailure = message;
         }
         const data = this.parseFfmpegInputReport(report);
-        return data
-            ? { success: true, data }
-            : { success: false, error: "ffmpeg reported no duration" };
+        if (data) {
+            return { success: true, data };
+        }
+        // Keep ffmpeg's own words when it failed as well as found no duration:
+        // a file it cannot read at all is the common case, and the reason is the
+        // only thing the warning that follows can tell the reader.
+        return {
+            success: false,
+            error: exitFailure
+                ? `ffmpeg reported no duration: ${exitFailure}`
+                : "ffmpeg reported no duration",
+        };
     }
     /**
      * Parse the input section of an ffmpeg banner: `Duration:`, `bitrate:` and