@juspay/neurolink 9.94.5 → 9.94.6

package/dist/lib/utils/messageBuilder.js

@@ -5,12 +5,12 @@ import { MultimodalLogger, ProviderImageAdapter, } from "../adapters/providerIma
 import { CONVERSATION_INSTRUCTIONS, STRUCTURED_OUTPUT_INSTRUCTIONS, } from "../config/conversationMemory.js";
 import { getAvailableInputTokens } from "../constants/contextWindows.js";
 import { enforceAggregateFileBudget, FILE_READ_BUDGET_PERCENT, } from "../context/fileTokenBudget.js";
-import { SIZE_TIER_THRESHOLDS } from "../types/index.js";
+import { isCSVContent, SIZE_TIER_THRESHOLDS } from "../types/index.js";
 import { tracers, ATTR, withSpan } from "../telemetry/index.js";
-import { NeuroLinkError } from "./errorHandling.js";
+import { ErrorFactory, NeuroLinkError, withTimeout } from "./errorHandling.js";
 import { FileDetector } from "./fileDetector.js";
 import { getImageCache } from "./imageCache.js";
-import { ImageProcessor } from "./imageProcessor.js";
+import { ImageProcessor, imageUtils } from "./imageProcessor.js";
 import { logger } from "./logger.js";
 import { PDFImageConverter, PDFProcessor } from "./pdfProcessor.js";
 import { urlDownloadRateLimiter } from "./rateLimiter.js";
@@ -860,7 +860,11 @@ export async function processUnifiedFilesArray(options, maxSize, provider) {
             }
             catch (error) {
                 const errMsg = error instanceof Error ? error.message : String(error);
-                logger.error(`[NEUROLINK] File skipped/failed: ${filename} — reason: ${errMsg}`);
+                // #273: don't silently drop a failed file — log, then throw so the
+                // caller learns the file couldn't be processed (matches the explicit
+                // pdf/csv paths' fail-loud behavior).
+                logger.error(`[NEUROLINK] File processing failed: ${filename} — reason: ${errMsg}`);
+                throw ErrorFactory.fileProcessingFailed(filename, error instanceof Error ? error : new Error(errMsg));
             }
         }
         span.setAttribute(ATTR.FILE_INCLUDED_COUNT, includedCount);
@@ -935,10 +939,12 @@ async function processExplicitCsvFiles(options) {
             logger.info(`[CSV] ✅ Processed: ${filename}`);
         }
         catch (error) {
-            logger.error(`[CSV] ❌ Failed:`, error);
             const filename = extractFilename(csvFile, i);
-            options.input.text += `\n\n## CSV Data Error: Failed to process "${filename}"`;
-            options.input.text += `\nReason: ${error instanceof Error ? error.message : "Unknown error"}`;
+            const errMsg = error instanceof Error ? error.message : String(error);
+            // #273: fail loud instead of embedding the error into the prompt text
+            // (which the model would then "analyze"). Log, then throw.
+            logger.error(`[CSV] ❌ Failed to process ${filename}: ${errMsg}`);
+            throw ErrorFactory.csvProcessingFailed(filename, error instanceof Error ? error : new Error(errMsg));
         }
     }
 }
@@ -1067,6 +1073,18 @@ export async function buildMultimodalMessagesArray(options, provider, model) {
     // local const so TypeScript sees the definite (non-optional) type in the
     // rest of this function, avoiding 60+ "possibly undefined" errors.
     const inp = options.input;
+    // #284: audioFiles/videoFiles have no dedicated processor — route them
+    // through the same auto-detecting `files` pipeline that already
+    // understands "audio"/"video" FileDetector results (see
+    // appendDetectedFileResult), instead of silently dropping them once
+    // detectMultimodal() routes an audio/video-only request here.
+    if (inp.audioFiles?.length || inp.videoFiles?.length) {
+        inp.files = [
+            ...(inp.files || []),
+            ...(inp.audioFiles || []),
+            ...(inp.videoFiles || []),
+        ];
+    }
     // Compute provider-specific max PDF size once for consistent validation
     const pdfConfig = PDFProcessor.getProviderConfig(provider);
     const maxSize = pdfConfig
@@ -1088,6 +1106,13 @@ export async function buildMultimodalMessagesArray(options, provider, model) {
     const hasPDFs = pdfFiles.length > 0;
     // If no images or PDFs, use standard message building and convert to MultimodalChatMessage[]
     if (!hasImages && !hasPDFs) {
+        // #289: CSV content[] items don't need vision, so they never reach the
+        // multimodal converter below — process them into the prompt text here
+        // (otherwise a `content: [{type:"csv"}]`-only request silently drops it).
+        const csvContentItems = inp.content?.filter(isCSVContent) ?? [];
+        if (csvContentItems.length > 0) {
+            inp.text = await appendCsvContentToText(csvContentItems, inp.text ?? "");
+        }
         if (inp.csvFiles) {
             inp.csvFiles = [];
         }
@@ -1216,6 +1241,47 @@ export async function buildMultimodalMessagesArray(options, provider, model) {
         throw error;
     }
 }
+/**
+ * Timeout for detecting/parsing an in-memory CSV `content[]` buffer (#325
+ * review, round 2). Bounds a stalled detector rather than blocking the
+ * request indefinitely.
+ */
+const CSV_CONTENT_DETECTION_TIMEOUT_MS = 30_000;
+/**
+ * #289: process CSV `content[]` items into appended prompt text. CSV is
+ * delivered to the model as text (like the explicit `csvFiles` path), so this
+ * is shared by both the no-vision gate and the multimodal converter.
+ */
+async function appendCsvContentToText(csvItems, baseText) {
+    let text = baseText;
+    for (const csv of csvItems) {
+        const raw = csv.data;
+        if (raw === undefined) {
+            continue;
+        }
+        // #325: raw CSV text (e.g. "a,b\n1,2") is not base64 — decoding it as
+        // base64 silently corrupts the content. Only treat the string as base64
+        // when it actually validates as such; otherwise treat it as literal
+        // UTF-8 CSV text, matching how Buffer inputs are already handled as-is.
+        const buffer = typeof raw === "string"
+            ? imageUtils.isValidBase64(raw)
+                ? Buffer.from(raw, "base64")
+                : Buffer.from(raw, "utf-8")
+            : raw;
+        const name = csv.metadata?.filename || "data.csv";
+        // #325 review (round 2): a stalled/hung detector must not block the
+        // request indefinitely — wrap with the project's standard withTimeout.
+        const result = await withTimeout(FileDetector.detectAndProcess(buffer, {
+            allowedTypes: ["csv"],
+            csvOptions: {
+                maxRows: csv.metadata?.maxRows,
+                formatStyle: csv.metadata?.formatStyle,
+            },
+        }), CSV_CONTENT_DETECTION_TIMEOUT_MS, new Error(`Timed out processing CSV content "${name}" after ${CSV_CONTENT_DETECTION_TIMEOUT_MS}ms`));
+        text += `${text ? "\n\n" : ""}## CSV Data from ${name}\n${result.content}`;
+    }
+    return text;
+}
 /**
  * Convert advanced content format to provider-specific format
  */
@@ -1223,14 +1289,19 @@ async function convertContentToProviderFormat(content, provider, _model) {
     const textContent = content.find((c) => c.type === "text");
     const imageContent = content.filter((c) => c.type === "image");
     const pdfContent = content.filter((c) => c.type === "pdf");
+    const csvContent = content.filter(isCSVContent);
     // Allow empty text when multimodal content is present (enables image-only or PDF-only queries)
-    const text = textContent?.text || "";
+    let text = textContent?.text || "";
+    // #289: CSV content[] items were silently dropped — fold each into the text.
+    if (csvContent.length > 0) {
+        text = await appendCsvContentToText(csvContent, text);
+    }
     const hasMultimodal = imageContent.length > 0 || pdfContent.length > 0;
     // Validate that we have at least some content
     if (!hasMultimodal && !text) {
         throw new Error("Content must include either text or multimodal content");
     }
-    // Text-only case
+    // Text-only case (CSV has already been folded into `text`)
     if (imageContent.length === 0 && pdfContent.length === 0) {
         return text;
     }

package/dist/types/generate.d.ts

@@ -53,6 +53,7 @@ export type GenerateOptions = {
         images?: Array<Buffer | string | ImageWithAltText>;
         csvFiles?: Array<Buffer | string>;
         pdfFiles?: Array<Buffer | string>;
+        audioFiles?: Array<Buffer | string>;
         videoFiles?: Array<Buffer | string>;
         files?: Array<Buffer | string | FileWithMetadata>;
         content?: Content[];

package/dist/types/multimodal.d.ts

@@ -419,15 +419,46 @@ export type MultimodalInput = {
     segments?: DirectorSegment[];
 };
 /**
- * Content format for multimodal messages (used internally)
- * Compatible with Vercel AI SDK message format
+ * Content format for multimodal messages (used internally).
+ *
+ * #325: the loose `[key: string]: unknown` index signature has been replaced
+ * with the concrete fields the codebase actually reads/writes across the
+ * text / image / file / tool-call / tool-result shapes. This keeps the broad
+ * structural compatibility the internal pipeline relies on (a single object
+ * type, not a strict discriminated union that would force narrowing at every
+ * consumer) while removing the "any key is allowed" hole that let typos and
+ * unrelated keys through unchecked.
  */
 export type MessageContent = {
     type: string;
+    /** Text content (`type: "text"`). */
     text?: string;
+    /** Base64 / data-URI image (`type: "image"`). */
     image?: string;
+    /** MIME type for image/file parts. */
     mimeType?: string;
-    [key: string]: unknown;
+    /** Raw file bytes or base64 (`type: "file"`/document parts). */
+    data?: string | Buffer;
+    /** File name for document/file parts. */
+    name?: string;
+    /** File name (alias used by some file parts). */
+    filename?: string;
+    /** Tool-call identifier (`type: "tool-call"`/`"tool-result"`). */
+    toolCallId?: string;
+    /** Tool name (`type: "tool-call"`/`"tool-result"`). */
+    toolName?: string;
+    /** Tool-call arguments (`type: "tool-call"`). */
+    args?: Record<string, unknown>;
+    /** Tool-result payload (`type: "tool-result"`). */
+    result?: unknown;
+    /** Whether a tool-result represents an error (`type: "tool-result"`). */
+    isError?: boolean;
+    /**
+     * Provider-specific per-block options (e.g. Anthropic cache_control).
+     * Read as `item.providerOptions` when converting `MessageContent[]` to
+     * `ModelMessage[]` in `MessageBuilder.ts`.
+     */
+    providerOptions?: Record<string, unknown>;
 };
 /**
  * Extended chat message for multimodal support (internal use)

package/dist/types/stream.d.ts

@@ -193,6 +193,7 @@ export type StreamOptions = {
         images?: Array<Buffer | string | ImageWithAltText>;
         csvFiles?: Array<Buffer | string>;
         pdfFiles?: Array<Buffer | string>;
+        audioFiles?: Array<Buffer | string>;
         videoFiles?: Array<Buffer | string>;
         files?: Array<Buffer | string | FileWithMetadata>;
         content?: Content[];

package/dist/utils/errorHandling.d.ts

@@ -35,6 +35,9 @@ export declare const ERROR_CODES: {
     readonly IMAGE_TOO_SMALL: "IMAGE_TOO_SMALL";
     readonly INVALID_IMAGE_FORMAT: "INVALID_IMAGE_FORMAT";
     readonly INVALID_IMAGE_SIZE: "INVALID_IMAGE_SIZE";
+    readonly IMAGE_BUFFER_INVALID: "IMAGE_BUFFER_INVALID";
+    readonly FILE_PROCESSING_FAILED: "FILE_PROCESSING_FAILED";
+    readonly CSV_PROCESSING_FAILED: "CSV_PROCESSING_FAILED";
     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";
@@ -216,6 +219,13 @@ export declare class ErrorFactory {
      * Create an invalid image format error
      */
     static invalidImageFormat(): NeuroLinkError;
+    /**
+     * Create an image buffer validation error: empty, undersized, or a
+     * truncated buffer detected by `ImageProcessor.validateBufferNotEmpty()`.
+     * Takes the fully-formed message so call sites keep their specific
+     * byte-count detail instead of a fixed generic message.
+     */
+    static imageBufferInvalid(message: string): NeuroLinkError;
     /**
      * Create a rate limiter queue full error
      */
@@ -287,6 +297,18 @@ export declare class ErrorFactory {
      * generic `Error` instances.
      */
     static proxyWorkerLifecycle(message: string, context?: Record<string, unknown>): NeuroLinkError;
+    /**
+     * Create a generic file-processing-failed error (e.g. an unrecognized or
+     * corrupt file in `processUnifiedFilesArray`). Preserves the original error
+     * as `originalError` (stack + message copied onto the new error's context).
+     */
+    static fileProcessingFailed(filename: string, originalError: Error): NeuroLinkError;
+    /**
+     * Create a generic CSV-processing-failed error (explicit `csvFiles` path,
+     * distinct from `fileProcessingFailed` so callers can classify CSV-specific
+     * failures separately). Preserves the original error as `originalError`.
+     */
+    static csvProcessingFailed(filename: string, originalError: Error): NeuroLinkError;
 }
 /**
  * Timeout wrapper for async operations

package/dist/utils/errorHandling.js

@@ -46,6 +46,10 @@ export const ERROR_CODES = {
     IMAGE_TOO_SMALL: "IMAGE_TOO_SMALL",
     INVALID_IMAGE_FORMAT: "INVALID_IMAGE_FORMAT",
     INVALID_IMAGE_SIZE: "INVALID_IMAGE_SIZE",
+    IMAGE_BUFFER_INVALID: "IMAGE_BUFFER_INVALID",
+    // Generic file/CSV processing errors
+    FILE_PROCESSING_FAILED: "FILE_PROCESSING_FAILED",
+    CSV_PROCESSING_FAILED: "CSV_PROCESSING_FAILED",
     // PDF validation errors
     PDF_PAGE_LIMIT_EXCEEDED: "PDF_PAGE_LIMIT_EXCEEDED",
     // Rate limiter errors
@@ -627,6 +631,22 @@ export class ErrorFactory {
             },
         });
     }
+    /**
+     * Create an image buffer validation error: empty, undersized, or a
+     * truncated buffer detected by `ImageProcessor.validateBufferNotEmpty()`.
+     * Takes the fully-formed message so call sites keep their specific
+     * byte-count detail instead of a fixed generic message.
+     */
+    static imageBufferInvalid(message) {
+        return new NeuroLinkError({
+            code: ERROR_CODES.IMAGE_BUFFER_INVALID,
+            message,
+            category: ErrorCategory.VALIDATION,
+            severity: ErrorSeverity.MEDIUM,
+            retriable: false,
+            context: { field: "input.images" },
+        });
+    }
     // ============================================================================
     // RATE LIMITER ERRORS
     // ============================================================================
@@ -924,6 +944,41 @@ export class ErrorFactory {
             context: context || {},
         });
     }
+    // ============================================================================
+    // GENERIC FILE / CSV PROCESSING ERRORS
+    // ============================================================================
+    /**
+     * Create a generic file-processing-failed error (e.g. an unrecognized or
+     * corrupt file in `processUnifiedFilesArray`). Preserves the original error
+     * as `originalError` (stack + message copied onto the new error's context).
+     */
+    static fileProcessingFailed(filename, originalError) {
+        return new NeuroLinkError({
+            code: ERROR_CODES.FILE_PROCESSING_FAILED,
+            message: `Failed to process file "${filename}": ${originalError.message}`,
+            category: ErrorCategory.EXECUTION,
+            severity: ErrorSeverity.HIGH,
+            retriable: false,
+            context: { filename },
+            originalError,
+        });
+    }
+    /**
+     * Create a generic CSV-processing-failed error (explicit `csvFiles` path,
+     * distinct from `fileProcessingFailed` so callers can classify CSV-specific
+     * failures separately). Preserves the original error as `originalError`.
+     */
+    static csvProcessingFailed(filename, originalError) {
+        return new NeuroLinkError({
+            code: ERROR_CODES.CSV_PROCESSING_FAILED,
+            message: `Failed to process CSV file "${filename}": ${originalError.message}`,
+            category: ErrorCategory.EXECUTION,
+            severity: ErrorSeverity.HIGH,
+            retriable: false,
+            context: { filename },
+            originalError,
+        });
+    }
 }
 /**
  * Timeout wrapper for async operations

package/dist/utils/fileDetector.js

@@ -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 { withTimeout } from "./errorHandling.js";
 import { redactUrlForError, sanitizeErrorCause } from "./logSanitize.js";
 import { mimeHintToExtension, mimeHintToFileType, normalizeMimeHint, } from "./mimeTypeHints.js";
 import { PDFProcessor } from "./pdfProcessor.js";
@@ -32,6 +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)
  */
@@ -1351,6 +1426,9 @@ export class FileDetector {
                     // 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) {
@@ -1698,13 +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/imageProcessor.d.ts

@@ -4,6 +4,17 @@
  */
 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
  */
@@ -17,6 +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:

package/dist/utils/imageProcessor.js

@@ -11,6 +11,23 @@ import { SYSTEM_LIMITS } from "../core/constants.js";
 import { SIZE_LIMITS_BYTES } from "../processors/config/sizeLimits.js";
 import { getImageCache } from "./imageCache.js";
 import { ErrorFactory, NeuroLinkError } from "./errorHandling.js";
+/**
+ * Minimum bytes required just to detect an image format from magic bytes (#293).
+ */
+export 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 const MIN_VALID_IMAGE_SIZE = {
+    "image/png": 67,
+    "image/jpeg": 125,
+    "image/gif": 43,
+    "image/webp": 20,
+    "image/avif": 100,
+};
 /**
  * Network error codes that should trigger a retry
  */
@@ -92,12 +109,9 @@ export class ImageProcessor {
      * @returns Processed image as data URI
      */
     static async process(content, _options) {
-        // Validate content is non-empty before processing
-        if (content.length === 0) {
-            logger.error("Empty buffer provided");
-            throw new Error("Invalid image processing: buffer is empty");
-        }
-        const mediaType = this.detectImageType(content);
+        // #293: reject empty AND small/truncated buffers (not just 0 bytes) before
+        // processing; returns the detected media type.
+        const mediaType = this.validateBufferNotEmpty(content);
         // #261/#286 follow-up: fail loud instead of packaging the octet-stream
         // sentinel as a "valid" image (see assertKnownImageType() above).
         assertKnownImageType(mediaType);
@@ -115,6 +129,73 @@ export class ImageProcessor {
             },
         };
     }
+    /**
+     * 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).
+     */
+    static hasStrictImageMagic(content, mediaType) {
+        const b = content;
+        switch (mediaType) {
+            case "image/png":
+                return (b.length >= 4 &&
+                    b[0] === 0x89 &&
+                    b[1] === 0x50 &&
+                    b[2] === 0x4e &&
+                    b[3] === 0x47);
+            case "image/jpeg":
+                return b.length >= 3 && b[0] === 0xff && b[1] === 0xd8 && b[2] === 0xff;
+            case "image/gif":
+                return b.length >= 3 && b[0] === 0x47 && b[1] === 0x49 && b[2] === 0x46;
+            case "image/webp":
+                return (b.length >= 12 &&
+                    b.subarray(0, 4).toString("latin1") === "RIFF" &&
+                    b.subarray(8, 12).toString("latin1") === "WEBP");
+            case "image/avif":
+                return (b.length >= 12 &&
+                    b.subarray(4, 8).toString("latin1") === "ftyp" &&
+                    b.subarray(8, 12).toString("latin1").startsWith("avif"));
+            default:
+                return false;
+        }
+    }
+    /**
+     * 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.
+     */
+    static validateBufferNotEmpty(content) {
+        if (content.length === 0) {
+            logger.error("Empty buffer provided");
+            throw ErrorFactory.imageBufferInvalid("Invalid image processing: buffer is empty");
+        }
+        const mediaType = this.detectImageType(content);
+        // SVG is text and can be legitimately tiny (e.g. `<svg/>`), so the raster
+        // size floors don't apply to it (#293 review).
+        if (mediaType === "image/svg+xml") {
+            return mediaType;
+        }
+        // A buffer too small to carry any raster header is not a usable image.
+        if (content.length < MIN_IMAGE_SIZE) {
+            throw ErrorFactory.imageBufferInvalid(`Invalid image processing: buffer too small (${content.length} bytes) — ` +
+                `a minimum of ${MIN_IMAGE_SIZE} bytes is required for a valid image.`);
+        }
+        // Only enforce a per-format minimum when the format was strictly identified
+        // from magic bytes (never from detectImageType's jpeg fallback), so valid
+        // BMP/TIFF/unknown buffers are not mis-rejected as "truncated jpeg".
+        const minForFormat = MIN_VALID_IMAGE_SIZE[mediaType];
+        if (minForFormat !== undefined &&
+            content.length < minForFormat &&
+            this.hasStrictImageMagic(content, mediaType)) {
+            throw ErrorFactory.imageBufferInvalid(`Invalid image processing: ${mediaType} buffer is truncated ` +
+                `(${content.length} bytes, minimum ${minForFormat} for a valid ${mediaType}).`);
+        }
+        return mediaType;
+    }
     /**
      * Validate processed output meets required format
      * Checks: