@juspay/neurolink 9.94.3 → 9.94.5

package/dist/lib/utils/errorHandling.d.ts

@@ -34,6 +34,7 @@ export declare const ERROR_CODES: {
     readonly IMAGE_TOO_LARGE: "IMAGE_TOO_LARGE";
     readonly IMAGE_TOO_SMALL: "IMAGE_TOO_SMALL";
     readonly INVALID_IMAGE_FORMAT: "INVALID_IMAGE_FORMAT";
+    readonly INVALID_IMAGE_SIZE: "INVALID_IMAGE_SIZE";
     readonly PDF_PAGE_LIMIT_EXCEEDED: "PDF_PAGE_LIMIT_EXCEEDED";
     readonly RATE_LIMITER_QUEUE_FULL: "RATE_LIMITER_QUEUE_FULL";
     readonly RATE_LIMITER_QUEUE_TIMEOUT: "RATE_LIMITER_QUEUE_TIMEOUT";
@@ -199,6 +200,14 @@ export declare class ErrorFactory {
      * Create an image too large error
      */
     static imageTooLarge(sizeMB: string, maxMB: string): NeuroLinkError;
+    /**
+     * Create an invalid image size error (NaN/Infinity/negative byte length).
+     * Distinct from `imageTooLarge` — this rejects a malformed size value
+     * before it reaches the max-size comparison, since `NaN > maxSize` and
+     * `-1 > maxSize` both evaluate to `false` and would otherwise let a
+     * corrupted stat/header value silently skip the guard.
+     */
+    static invalidImageSize(size: number): NeuroLinkError;
     /**
      * Create an image too small error
      */

package/dist/lib/utils/errorHandling.js

@@ -45,6 +45,7 @@ export const ERROR_CODES = {
     IMAGE_TOO_LARGE: "IMAGE_TOO_LARGE",
     IMAGE_TOO_SMALL: "IMAGE_TOO_SMALL",
     INVALID_IMAGE_FORMAT: "INVALID_IMAGE_FORMAT",
+    INVALID_IMAGE_SIZE: "INVALID_IMAGE_SIZE",
     // PDF validation errors
     PDF_PAGE_LIMIT_EXCEEDED: "PDF_PAGE_LIMIT_EXCEEDED",
     // Rate limiter errors
@@ -570,6 +571,26 @@ export class ErrorFactory {
             },
         });
     }
+    /**
+     * Create an invalid image size error (NaN/Infinity/negative byte length).
+     * Distinct from `imageTooLarge` — this rejects a malformed size value
+     * before it reaches the max-size comparison, since `NaN > maxSize` and
+     * `-1 > maxSize` both evaluate to `false` and would otherwise let a
+     * corrupted stat/header value silently skip the guard.
+     */
+    static invalidImageSize(size) {
+        return new NeuroLinkError({
+            code: ERROR_CODES.INVALID_IMAGE_SIZE,
+            message: `Invalid image size: ${size} (must be a finite, non-negative number)`,
+            category: ErrorCategory.VALIDATION,
+            severity: ErrorSeverity.MEDIUM,
+            retriable: false,
+            context: {
+                field: "input.images",
+                size,
+            },
+        });
+    }
     /**
      * Create an image too small error
      */

package/dist/lib/utils/fileDetector.js

@@ -4,7 +4,7 @@
  * Uses multi-strategy approach for reliable type identification
  */
 import { open, readFile, realpath } from "fs/promises";
-import { isAbsolute as isAbsolutePath, relative as relativePath, resolve as resolvePath, sep, } from "path";
+import { basename, isAbsolute as isAbsolutePath, relative as relativePath, resolve as resolvePath, sep, } from "path";
 import { getGlobalDispatcher, interceptors, request } from "undici";
 // Lazy-loaded processor singletons — avoids loading heavy media deps
 // (mediabunny, fluent-ffmpeg, music-metadata, adm-zip) on every generate() call.
@@ -24,6 +24,7 @@ import { tracers, ATTR, withSpan } from "../telemetry/index.js";
 import { CSVProcessor } from "./csvProcessor.js";
 import { ImageProcessor } from "./imageProcessor.js";
 import { logger } from "./logger.js";
+import { redactUrlForError, sanitizeErrorCause } from "./logSanitize.js";
 import { mimeHintToExtension, mimeHintToFileType, normalizeMimeHint, } from "./mimeTypeHints.js";
 import { PDFProcessor } from "./pdfProcessor.js";
 /**
@@ -1338,25 +1339,55 @@ export class FileDetector {
         const maxRetries = options?.maxRetries ?? DEFAULT_MAX_RETRIES;
         const retryDelay = options?.retryDelay ?? DEFAULT_RETRY_DELAY;
         return withRetry(async () => {
-            const response = await request(url, {
-                dispatcher: getGlobalDispatcher().compose(interceptors.redirect({ maxRedirections: 5 })),
-                method: "GET",
-                headersTimeout: timeout,
-                bodyTimeout: timeout,
-            });
-            if (response.statusCode !== 200) {
-                throw new Error(`HTTP ${response.statusCode} fetching ${url}`);
-            }
-            const chunks = [];
-            let totalSize = 0;
-            for await (const chunk of response.body) {
-                totalSize += chunk.length;
-                if (totalSize > maxSize) {
-                    throw new Error(`File too large: ${formatFileSize(totalSize)} (max: ${formatFileSize(maxSize)})`);
+            try {
+                const response = await request(url, {
+                    dispatcher: getGlobalDispatcher().compose(interceptors.redirect({ maxRedirections: 5 })),
+                    method: "GET",
+                    headersTimeout: timeout,
+                    bodyTimeout: timeout,
+                });
+                if (response.statusCode !== 200) {
+                    // Query string / fragment stripped — a presigned URL's token must
+                    // not be echoed into a thrown error.
+                    throw new Error(`HTTP ${response.statusCode} fetching ${redactUrlForError(url)}`);
                 }
-                chunks.push(chunk);
+                const chunks = [];
+                let totalSize = 0;
+                for await (const chunk of response.body) {
+                    totalSize += chunk.length;
+                    if (totalSize > maxSize) {
+                        throw new Error(`File too large: ${formatFileSize(totalSize)} (max: ${formatFileSize(maxSize)})`);
+                    }
+                    chunks.push(chunk);
+                }
+                return Buffer.concat(chunks);
+            }
+            catch (error) {
+                // Node/undici DNS, TLS, and connect-timeout errors embed the full
+                // request URL (including a presigned query token) in
+                // `error.message`. Redact into a NEW error instead of mutating the
+                // original in place, so anything that still holds a reference to
+                // the original — debug logs, telemetry spans, upstream callers —
+                // keeps seeing the real message. `.code` is copied onto the new
+                // error so `isRetryableNetworkError`'s retry check in the outer
+                // `withRetry` catch still classifies it correctly. The raw
+                // original error is NEVER attached as `cause` — that would leave
+                // the unredacted URL reachable via `error.cause.message` for
+                // anything that walks the cause chain (cause-aware logging,
+                // telemetry). `cause` instead gets its own sanitized copy.
+                // `sanitizeErrorCause` handles non-`Error` thrown values too (a raw
+                // string/object can just as easily carry the full URL), so there is
+                // no unconditional `throw error` fallback that would bypass
+                // redaction for that case.
+                const cause = sanitizeErrorCause(error);
+                const redacted = new Error(cause.message, { cause });
+                redacted.name = cause.name;
+                const code = cause.code;
+                if (code !== undefined) {
+                    redacted.code = code;
+                }
+                throw redacted;
             }
-            return Buffer.concat(chunks);
         }, { maxRetries, retryDelay });
     }
     /**
@@ -1374,6 +1405,13 @@ export class FileDetector {
         // the base dir pointing outside cannot bypass containment. The
         // path.relative check (not a string prefix) correctly handles the root dir
         // ("/") and sibling-prefix ("/app" vs "/app-evil") edge cases.
+        //
+        // The actual open() below MUST target this validated `real` path, not the
+        // original `filePath` — otherwise a symlink swapped between the realpath()
+        // check and the open() call routes the read outside the sandbox even
+        // though validation passed (TOCTOU). With no sandbox configured there's no
+        // boundary to defend, so the original path is used as given.
+        let pathToOpen = filePath;
         if (options?.allowedBaseDir) {
             let base;
             let real;
@@ -1381,24 +1419,66 @@ export class FileDetector {
                 base = await realpath(resolvePath(options.allowedBaseDir));
                 real = await realpath(filePath);
             }
-            catch {
-                throw new Error(`Access denied: "${filePath}" could not be resolved within the allowed base directory`);
+            catch (error) {
+                // Full path stays in the debug log; the thrown (potentially
+                // client-facing) error only gets the basename to avoid leaking the
+                // host's directory layout to an untrusted caller. The cause is
+                // sanitized too — Node's realpath ENOENT/EACCES messages embed the
+                // full path verbatim, which would otherwise survive on the cause
+                // chain (cause-aware logging, telemetry) even though the outer
+                // message is already redacted.
+                logger.debug("loadFromPath: realpath resolution failed", {
+                    filePath,
+                    error,
+                });
+                // Assigned to a variable before the throw (rather than an inline
+                // `{ cause: sanitizeErrorCause(...) }`) so the sanitized, path-redacted
+                // copy is unambiguously the attached cause — the raw `error`, whose
+                // message still embeds the full path, is never reachable from the
+                // thrown result.
+                const cause = sanitizeErrorCause(error, { filePath });
+                const denied = new Error(`Access denied: "${basename(filePath)}" could not be resolved within the allowed base directory`, { cause });
+                throw denied;
             }
             const rel = relativePath(base, real);
             if (rel === ".." || rel.startsWith(`..${sep}`) || isAbsolutePath(rel)) {
-                throw new Error(`Access denied: "${filePath}" resolves outside the allowed base directory`);
+                logger.debug("loadFromPath: path resolves outside allowed base dir", {
+                    filePath,
+                    real,
+                });
+                throw new Error(`Access denied: "${basename(filePath)}" resolves outside the allowed base directory`);
             }
+            pathToOpen = real;
         }
         // Open a handle and stat/read through the SAME descriptor so a symlink
         // swap between the size check and the read cannot occur (TOCTOU).
-        const handle = await open(filePath, "r");
+        let handle;
+        try {
+            handle = await open(pathToOpen, "r");
+        }
+        catch (error) {
+            // A failed open (ENOENT/EACCES/…) embeds the opened path verbatim in
+            // its message. When a sandbox is configured `pathToOpen` is the
+            // realpath-resolved target (`real`), which differs from both `filePath`
+            // and its resolved form — so redact `pathToOpen` specifically, or the
+            // full host path would survive on both the thrown message and the cause
+            // chain despite this PR's path-redaction hardening.
+            const cause = sanitizeErrorCause(error, { filePath: pathToOpen });
+            const failed = new Error(cause.message, { cause });
+            failed.name = cause.name;
+            const code = cause.code;
+            if (code !== undefined) {
+                failed.code = code;
+            }
+            throw failed;
+        }
         try {
             const statInfo = await handle.stat();
             if (!statInfo.isFile()) {
-                throw new Error(`Not a file: ${filePath}`);
+                throw new Error(`Not a file: ${basename(filePath)}`);
             }
             if (statInfo.size > maxSize) {
-                throw new Error(`File too large: ${filePath} is ${formatFileSize(statInfo.size)} (max: ${formatFileSize(maxSize)})`);
+                throw new Error(`File too large: ${basename(filePath)} is ${formatFileSize(statInfo.size)} (max: ${formatFileSize(maxSize)})`);
             }
             return await handle.readFile();
         }
@@ -1450,6 +1530,15 @@ class MagicBytesStrategy {
             input[6] === 0x79 &&
             input[7] === 0x70) {
             const brand = input.length >= 12 ? input.toString("latin1", 8, 12) : "";
+            // AVIF images share the ISO-BMFF ftyp box with MP4/MOV; the major brand
+            // ('avif' still, 'avis' sequence, 'avio' intra-only AV1 image/sequence
+            // — spec-listed under compatible_brands but also emitted as
+            // major_brand by real encoders) distinguishes them. Detect before the
+            // audio/video branches so an AVIF buffer isn't misrouted to the video
+            // pipeline (#286).
+            if (/^(avif|avis|avio)/.test(brand)) {
+                return this.result("image", "image/avif", 95);
+            }
             if (/^(M4A|M4B|M4P|F4A|F4B)/.test(brand)) {
                 return this.result("audio", "audio/mp4", 95);
             }

package/dist/lib/utils/imageCache.js

@@ -14,6 +14,7 @@
  */
 import { createHash } from "crypto";
 import { logger } from "./logger.js";
+import { redactUrlForError } from "./logSanitize.js";
 /**
  * LRU Cache for downloaded images
  *
@@ -142,7 +143,9 @@ export class ImageCache {
         const entry = this.cache.get(normalizedUrl);
         if (!entry) {
             this.stats.misses++;
-            logger.debug("Image cache miss", { url: normalizedUrl.substring(0, 50) });
+            logger.debug("Image cache miss", {
+                url: redactUrlForError(normalizedUrl),
+            });
             return null;
         }
         // Check TTL expiration
@@ -150,7 +153,7 @@ export class ImageCache {
             this.stats.expirations++;
             this.delete(normalizedUrl);
             logger.debug("Image cache entry expired", {
-                url: normalizedUrl.substring(0, 50),
+                url: redactUrlForError(normalizedUrl),
             });
             return null;
         }
@@ -162,7 +165,7 @@ export class ImageCache {
         this.cache.set(normalizedUrl, entry);
         this.stats.hits++;
         logger.debug("Image cache hit", {
-            url: normalizedUrl.substring(0, 50),
+            url: redactUrlForError(normalizedUrl),
             accessCount: entry.accessCount,
         });
         return entry;
@@ -192,7 +195,7 @@ export class ImageCache {
         // Skip caching if image exceeds max size
         if (size > this.maxImageSize) {
             logger.debug("Image too large to cache", {
-                url: normalizedUrl.substring(0, 50),
+                url: redactUrlForError(normalizedUrl),
                 size,
                 maxSize: this.maxImageSize,
             });
@@ -211,8 +214,8 @@ export class ImageCache {
                 // Update content hash index to point to the new URL as well
                 this.contentHashIndex.set(contentHash, normalizedUrl);
                 logger.debug("Image cache dedup hit", {
-                    newUrl: normalizedUrl.substring(0, 50),
-                    existingUrl: existingUrl.substring(0, 50),
+                    newUrl: redactUrlForError(normalizedUrl),
+                    existingUrl: redactUrlForError(existingUrl),
                 });
                 return;
             }
@@ -234,7 +237,7 @@ export class ImageCache {
         this.cache.set(normalizedUrl, entry);
         this.contentHashIndex.set(contentHash, normalizedUrl);
         logger.debug("Image cached", {
-            url: normalizedUrl.substring(0, 50),
+            url: redactUrlForError(normalizedUrl),
             size,
             contentHash: contentHash.substring(0, 8),
             cacheSize: this.cache.size,
@@ -272,7 +275,7 @@ export class ImageCache {
             this.cache.delete(oldestKey);
             this.stats.evictions++;
             logger.debug("Image cache eviction", {
-                url: String(oldestKey).substring(0, 50),
+                url: redactUrlForError(String(oldestKey)),
             });
         }
     }

package/dist/lib/utils/imageProcessor.d.ts

@@ -2,6 +2,7 @@
  * Image processing utilities for multimodal support
  * Handles format conversion for different AI providers
  */
+import { redactPathFromMessage } from "./logSanitize.js";
 import type { ProcessedImage, FileProcessingResult } from "../types/index.js";
 /**
  * Image processor class for handling provider-specific image formatting
@@ -59,9 +60,41 @@ export declare class ImageProcessor {
      */
     static detectImageType(input: string | Buffer): string;
     /**
-     * Validate image size (default 10MB limit)
+     * Throwing size guard for a raw byte length. Prevents memory exhaustion by
+     * rejecting an oversized input BEFORE an unbounded read/allocation happens
+     * (#257) — callers that can get a size cheaply (e.g. `fs.stat`) should
+     * validate it before ever reading the bytes into memory. Default limit is
+     * the canonical `SIZE_LIMITS_BYTES.IMAGE_MAX` (10 MB); there is no public
+     * way to raise it — `maxSize` is an internal-only override for advanced
+     * callers (e.g. video image inputs use a higher limit).
+     *
+     * @param size - Byte length to validate
+     * @param context - Short label identifying the source (unused in the
+     *   thrown message today, kept for call-site readability and future use)
+     * @param maxSize - Max allowed size in bytes
+     * @throws NeuroLinkError (`INVALID_IMAGE_SIZE`) if `size` is not a finite,
+     *   non-negative number
+     * @throws NeuroLinkError (`IMAGE_TOO_LARGE`) if `size` exceeds `maxSize`
+     */
+    static validateSize(size: number, context: string, maxSize?: number): void;
+    /**
+     * Throwing size guard for image buffers. Prevents memory exhaustion by
+     * rejecting oversized buffers with a descriptive error BEFORE an unbounded
+     * `Buffer.toString("base64")` allocation (#257). Delegates to
+     * `validateSize` so buffer-based and pre-read (stat-based) callers share
+     * one implementation.
+     *
+     * @param buffer - Image buffer to validate
+     * @param context - Short label identifying the source (see `validateSize`)
+     * @param maxSize - Max allowed size in bytes
+     * @throws NeuroLinkError (`IMAGE_TOO_LARGE`) if the buffer exceeds `maxSize`
+     */
+    static validateBufferSize(buffer: Buffer, context: string, maxSize?: number): void;
+    /**
+     * Size-guarded Buffer → base64. Use in place of `buffer.toString("base64")`
+     * on any path that converts a caller-supplied image buffer (#257).
      */
-    static validateImageSize(data: Buffer | string, maxSize?: number): boolean;
+    static safeBase64Convert(buffer: Buffer, context: string, maxSize?: number): string;
     /**
      * Validate image format
      */
@@ -126,13 +159,28 @@ export declare const imageUtils: {
      */
     formatFileSize: (bytes: number) => string;
     /**
-     * Convert Buffer to base64 string
+     * Convert Buffer to base64 string. Guarded by the same #257 size limit as
+     * every other conversion site (default `SIZE_LIMITS_BYTES.IMAGE_MAX`);
+     * pass `maxBytes` to override for a caller with a legitimate need for a
+     * different ceiling rather than silently accepting an unbounded buffer.
+     * `imageUtils`/`ImageProcessor` are internal-only — not re-exported from
+     * `src/lib/index.ts` or the package's public `exports` map — so this
+     * default does not change behavior for any external SDK consumer.
      */
-    bufferToBase64: (buffer: Buffer) => string;
+    bufferToBase64: (buffer: Buffer, maxBytes?: number) => string;
     /**
      * Convert base64 string to Buffer
      */
     base64ToBuffer: (base64: string) => Buffer;
+    /**
+     * Redact `filePath` (both as given and its `path.resolve()`'d form) from
+     * an error message. Exposed directly (not just used internally by
+     * {@link fileToBase64DataUri}) so the resolved-path branch can be verified
+     * by a deterministic unit test — real `fs` errors only ever embed the
+     * literal path as passed, never a resolved variant, so that branch isn't
+     * otherwise observable through the async file-reading API.
+     */
+    redactPathFromMessage: typeof redactPathFromMessage;
     /**
      * Convert file path to base64 data URI
      */