@juspay/neurolink 9.94.4 → 9.94.6

package/dist/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 +24,8 @@ import { tracers, ATTR, withSpan } from "../telemetry/index.js";
 import { CSVProcessor } from "./csvProcessor.js";
 import { ImageProcessor } from "./imageProcessor.js";
 import { logger } from "./logger.js";
+import { withTimeout } from "./errorHandling.js";
+import { redactUrlForError, sanitizeErrorCause } from "./logSanitize.js";
 import { mimeHintToExtension, mimeHintToFileType, normalizeMimeHint, } from "./mimeTypeHints.js";
 import { PDFProcessor } from "./pdfProcessor.js";
 /**
@@ -31,6 +33,80 @@ import { PDFProcessor } from "./pdfProcessor.js";
  */
 const DEFAULT_MAX_RETRIES = 3;
 const DEFAULT_RETRY_DELAY = 1000; // milliseconds
+/**
+ * Short-TTL cache of URL → Content-Type (#323). A URL is commonly detected more
+ * than once (repeated multimodal prompts reuse the same asset URL); caching the
+ * HEAD's content-type avoids re-issuing the HEAD each time. `loadFromURL` also
+ * populates it from its GET response, so once a URL's body has been fetched a
+ * subsequent detection needs no network round-trip at all.
+ *
+ * Trade-off (round-2 review): this is a module-level cache shared by every
+ * request in the process, and correctness relies solely on the 60s TTL — a
+ * signed URL whose response changes at the same path within that window would
+ * read stale. That is an intentional, bounded trade-off (60s of possible
+ * staleness for far fewer HEAD round-trips), not a freshness guarantee.
+ * Expired entries are removed lazily on their next `get()` (see
+ * `getCachedUrlContentType`); `setCachedUrlContentType` additionally sweeps
+ * expired entries opportunistically once the cache hits its size cap, so a
+ * URL that is cached once and never looked up again doesn't linger until the
+ * FIFO eviction below forces it out.
+ */
+const URL_CONTENT_TYPE_TTL_MS = 60_000;
+const URL_CONTENT_TYPE_CACHE_MAX_SIZE = 512;
+const urlContentTypeCache = new Map();
+function getCachedUrlContentType(url, now) {
+    const hit = urlContentTypeCache.get(url);
+    if (hit && hit.expiresAt > now) {
+        // Bump recency: Map iteration order follows insertion order, and the
+        // eviction below deletes the *first* key, so a plain `get` on a hot
+        // entry would leave it first in line for eviction despite being the
+        // most recently used. Re-inserting turns the size-bounded FIFO below
+        // into an actual LRU.
+        urlContentTypeCache.delete(url);
+        urlContentTypeCache.set(url, hit);
+        return hit.contentType;
+    }
+    if (hit) {
+        // Entry exists but its TTL has passed — treat as a miss and evict it
+        // immediately rather than serving (or retaining) stale data.
+        urlContentTypeCache.delete(url);
+    }
+    return undefined;
+}
+/**
+ * Opportunistically remove already-expired entries. Only called once the
+ * cache is at its size cap (see `setCachedUrlContentType`) so it doesn't add
+ * an O(n) scan to the common-case hot path.
+ */
+function pruneExpiredUrlContentTypeEntries(now) {
+    for (const [key, entry] of urlContentTypeCache) {
+        if (entry.expiresAt <= now) {
+            urlContentTypeCache.delete(key);
+        }
+    }
+}
+function setCachedUrlContentType(url, contentType, now) {
+    if (!contentType) {
+        return;
+    }
+    urlContentTypeCache.set(url, {
+        contentType,
+        expiresAt: now + URL_CONTENT_TYPE_TTL_MS,
+    });
+    // Bound the cache so a long-lived process can't grow it unbounded. Prefer
+    // reclaiming already-expired entries first; only fall back to evicting the
+    // oldest still-live entry (FIFO/LRU-ish, see getCachedUrlContentType) if
+    // the cache is still over the cap after pruning.
+    if (urlContentTypeCache.size > URL_CONTENT_TYPE_CACHE_MAX_SIZE) {
+        pruneExpiredUrlContentTypeEntries(now);
+    }
+    if (urlContentTypeCache.size > URL_CONTENT_TYPE_CACHE_MAX_SIZE) {
+        const oldest = urlContentTypeCache.keys().next().value;
+        if (oldest !== undefined) {
+            urlContentTypeCache.delete(oldest);
+        }
+    }
+}
 /**
  * Retryable network error codes (Node.js/undici network errors)
  */
@@ -1338,25 +1414,58 @@ 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)}`);
+                }
+                // #323: cache the Content-Type from this GET so a subsequent detection
+                // of the same URL needs no HEAD.
+                setCachedUrlContentType(url, response.headers["content-type"] || "", Date.now());
+                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;
                 }
-                chunks.push(chunk);
+                throw redacted;
             }
-            return Buffer.concat(chunks);
         }, { maxRetries, retryDelay });
     }
     /**
@@ -1374,6 +1483,13 @@ 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 +1497,66 @@ 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 +1608,15 @@ 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);
             }
@@ -1609,13 +1776,25 @@ class MimeTypeStrategy {
             return this.unknown();
         }
         try {
-            const response = await request(input, {
-                dispatcher: getGlobalDispatcher().compose(interceptors.redirect({ maxRedirections: 5 })),
-                method: "HEAD",
-                headersTimeout: FileDetector.DEFAULT_HEAD_TIMEOUT,
-                bodyTimeout: FileDetector.DEFAULT_HEAD_TIMEOUT,
-            });
-            const contentType = response.headers["content-type"] || "";
+            // #323: reuse a recently-seen Content-Type for this URL instead of
+            // re-issuing a HEAD (populated here and by loadFromURL's GET).
+            const now = Date.now();
+            let contentType = getCachedUrlContentType(input, now);
+            if (contentType === undefined) {
+                // Wrap the whole HEAD (request + body drain) in withTimeout so a stalled
+                // dump() can't hang detection, per the project's async-timeout guideline.
+                contentType = await withTimeout((async () => {
+                    const response = await request(input, {
+                        dispatcher: getGlobalDispatcher().compose(interceptors.redirect({ maxRedirections: 5 })),
+                        method: "HEAD",
+                        headersTimeout: FileDetector.DEFAULT_HEAD_TIMEOUT,
+                        bodyTimeout: FileDetector.DEFAULT_HEAD_TIMEOUT,
+                    });
+                    await response.body.dump();
+                    return response.headers["content-type"] || "";
+                })(), FileDetector.DEFAULT_HEAD_TIMEOUT);
+                setCachedUrlContentType(input, contentType, now);
+            }
             const type = this.mimeToFileType(contentType);
             return {
                 type,

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

@@ -2,7 +2,19 @@
  * 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";
+/**
+ * Minimum bytes required just to detect an image format from magic bytes (#293).
+ */
+export declare const MIN_IMAGE_SIZE = 12;
+/**
+ * Minimum plausible byte size for a non-empty image of each format (#293). A
+ * buffer that carries a format's magic bytes but is smaller than this is
+ * truncated/corrupt and would fail (or render blank) downstream — reject early
+ * with a clear message instead.
+ */
+export declare const MIN_VALID_IMAGE_SIZE: Record<string, number>;
 /**
  * Image processor class for handling provider-specific image formatting
  */
@@ -16,6 +28,22 @@ export declare class ImageProcessor {
      * @returns Processed image as data URI
      */
     static process(content: Buffer, _options?: unknown): Promise<FileProcessingResult>;
+    /**
+     * Strict magic-byte match: does `content` actually begin with `mediaType`'s
+     * signature? `detectImageType` returns `image/jpeg` as a *fallback* for
+     * unrecognized data, so a per-format minimum must only be enforced when the
+     * bytes genuinely match — otherwise valid BMP/TIFF/unknown buffers would be
+     * wrongly rejected as "truncated jpeg" (#293 review).
+     */
+    private static hasStrictImageMagic;
+    /**
+     * Validate that a buffer is a plausibly-complete image (#293): non-empty and,
+     * for a strictly-identified raster format, at least the minimum plausible byte
+     * size. Returns the detected media type.
+     *
+     * @throws Error when the buffer is empty or a recognized format is truncated.
+     */
+    private static validateBufferNotEmpty;
     /**
      * Validate processed output meets required format
      * Checks:
@@ -171,6 +199,15 @@ export declare const imageUtils: {
      * 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
      */