@juspay/neurolink 9.94.3 → 9.94.5
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +12 -0
- package/dist/browser/neurolink.min.js +381 -381
- package/dist/lib/utils/errorHandling.d.ts +9 -0
- package/dist/lib/utils/errorHandling.js +21 -0
- package/dist/lib/utils/fileDetector.js +113 -24
- package/dist/lib/utils/imageCache.js +11 -8
- package/dist/lib/utils/imageProcessor.d.ts +52 -4
- package/dist/lib/utils/imageProcessor.js +231 -29
- package/dist/lib/utils/logSanitize.d.ts +100 -1
- package/dist/lib/utils/logSanitize.js +152 -2
- package/dist/lib/utils/messageBuilder.js +49 -11
- package/dist/utils/errorHandling.d.ts +9 -0
- package/dist/utils/errorHandling.js +21 -0
- package/dist/utils/fileDetector.js +113 -24
- package/dist/utils/imageCache.js +11 -8
- package/dist/utils/imageProcessor.d.ts +52 -4
- package/dist/utils/imageProcessor.js +231 -29
- package/dist/utils/logSanitize.d.ts +100 -1
- package/dist/utils/logSanitize.js +152 -2
- package/dist/utils/messageBuilder.js +49 -11
- package/package.json +1 -1
|
@@ -2,6 +2,7 @@
|
|
|
2
2
|
* Image processing utilities for multimodal support
|
|
3
3
|
* Handles format conversion for different AI providers
|
|
4
4
|
*/
|
|
5
|
+
import { redactPathFromMessage } from "./logSanitize.js";
|
|
5
6
|
import type { ProcessedImage, FileProcessingResult } from "../types/index.js";
|
|
6
7
|
/**
|
|
7
8
|
* Image processor class for handling provider-specific image formatting
|
|
@@ -59,9 +60,41 @@ export declare class ImageProcessor {
|
|
|
59
60
|
*/
|
|
60
61
|
static detectImageType(input: string | Buffer): string;
|
|
61
62
|
/**
|
|
62
|
-
*
|
|
63
|
+
* Throwing size guard for a raw byte length. Prevents memory exhaustion by
|
|
64
|
+
* rejecting an oversized input BEFORE an unbounded read/allocation happens
|
|
65
|
+
* (#257) — callers that can get a size cheaply (e.g. `fs.stat`) should
|
|
66
|
+
* validate it before ever reading the bytes into memory. Default limit is
|
|
67
|
+
* the canonical `SIZE_LIMITS_BYTES.IMAGE_MAX` (10 MB); there is no public
|
|
68
|
+
* way to raise it — `maxSize` is an internal-only override for advanced
|
|
69
|
+
* callers (e.g. video image inputs use a higher limit).
|
|
70
|
+
*
|
|
71
|
+
* @param size - Byte length to validate
|
|
72
|
+
* @param context - Short label identifying the source (unused in the
|
|
73
|
+
* thrown message today, kept for call-site readability and future use)
|
|
74
|
+
* @param maxSize - Max allowed size in bytes
|
|
75
|
+
* @throws NeuroLinkError (`INVALID_IMAGE_SIZE`) if `size` is not a finite,
|
|
76
|
+
* non-negative number
|
|
77
|
+
* @throws NeuroLinkError (`IMAGE_TOO_LARGE`) if `size` exceeds `maxSize`
|
|
78
|
+
*/
|
|
79
|
+
static validateSize(size: number, context: string, maxSize?: number): void;
|
|
80
|
+
/**
|
|
81
|
+
* Throwing size guard for image buffers. Prevents memory exhaustion by
|
|
82
|
+
* rejecting oversized buffers with a descriptive error BEFORE an unbounded
|
|
83
|
+
* `Buffer.toString("base64")` allocation (#257). Delegates to
|
|
84
|
+
* `validateSize` so buffer-based and pre-read (stat-based) callers share
|
|
85
|
+
* one implementation.
|
|
86
|
+
*
|
|
87
|
+
* @param buffer - Image buffer to validate
|
|
88
|
+
* @param context - Short label identifying the source (see `validateSize`)
|
|
89
|
+
* @param maxSize - Max allowed size in bytes
|
|
90
|
+
* @throws NeuroLinkError (`IMAGE_TOO_LARGE`) if the buffer exceeds `maxSize`
|
|
91
|
+
*/
|
|
92
|
+
static validateBufferSize(buffer: Buffer, context: string, maxSize?: number): void;
|
|
93
|
+
/**
|
|
94
|
+
* Size-guarded Buffer → base64. Use in place of `buffer.toString("base64")`
|
|
95
|
+
* on any path that converts a caller-supplied image buffer (#257).
|
|
63
96
|
*/
|
|
64
|
-
static
|
|
97
|
+
static safeBase64Convert(buffer: Buffer, context: string, maxSize?: number): string;
|
|
65
98
|
/**
|
|
66
99
|
* Validate image format
|
|
67
100
|
*/
|
|
@@ -126,13 +159,28 @@ export declare const imageUtils: {
|
|
|
126
159
|
*/
|
|
127
160
|
formatFileSize: (bytes: number) => string;
|
|
128
161
|
/**
|
|
129
|
-
* Convert Buffer to base64 string
|
|
162
|
+
* Convert Buffer to base64 string. Guarded by the same #257 size limit as
|
|
163
|
+
* every other conversion site (default `SIZE_LIMITS_BYTES.IMAGE_MAX`);
|
|
164
|
+
* pass `maxBytes` to override for a caller with a legitimate need for a
|
|
165
|
+
* different ceiling rather than silently accepting an unbounded buffer.
|
|
166
|
+
* `imageUtils`/`ImageProcessor` are internal-only — not re-exported from
|
|
167
|
+
* `src/lib/index.ts` or the package's public `exports` map — so this
|
|
168
|
+
* default does not change behavior for any external SDK consumer.
|
|
130
169
|
*/
|
|
131
|
-
bufferToBase64: (buffer: Buffer) => string;
|
|
170
|
+
bufferToBase64: (buffer: Buffer, maxBytes?: number) => string;
|
|
132
171
|
/**
|
|
133
172
|
* Convert base64 string to Buffer
|
|
134
173
|
*/
|
|
135
174
|
base64ToBuffer: (base64: string) => Buffer;
|
|
175
|
+
/**
|
|
176
|
+
* Redact `filePath` (both as given and its `path.resolve()`'d form) from
|
|
177
|
+
* an error message. Exposed directly (not just used internally by
|
|
178
|
+
* {@link fileToBase64DataUri}) so the resolved-path branch can be verified
|
|
179
|
+
* by a deterministic unit test — real `fs` errors only ever embed the
|
|
180
|
+
* literal path as passed, never a resolved variant, so that branch isn't
|
|
181
|
+
* otherwise observable through the async file-reading API.
|
|
182
|
+
*/
|
|
183
|
+
redactPathFromMessage: typeof redactPathFromMessage;
|
|
136
184
|
/**
|
|
137
185
|
* Convert file path to base64 data URI
|
|
138
186
|
*/
|
|
@@ -2,11 +2,15 @@
|
|
|
2
2
|
* Image processing utilities for multimodal support
|
|
3
3
|
* Handles format conversion for different AI providers
|
|
4
4
|
*/
|
|
5
|
+
import { basename } from "path";
|
|
5
6
|
import { logger } from "./logger.js";
|
|
7
|
+
import { redactPathFromMessage, redactUrlForError, redactUrlsInText, sanitizeErrorCause, } from "./logSanitize.js";
|
|
6
8
|
import { urlDownloadRateLimiter } from "./rateLimiter.js";
|
|
7
9
|
import { withRetry } from "./retryHandler.js";
|
|
8
10
|
import { SYSTEM_LIMITS } from "../core/constants.js";
|
|
11
|
+
import { SIZE_LIMITS_BYTES } from "../processors/config/sizeLimits.js";
|
|
9
12
|
import { getImageCache } from "./imageCache.js";
|
|
13
|
+
import { ErrorFactory, NeuroLinkError } from "./errorHandling.js";
|
|
10
14
|
/**
|
|
11
15
|
* Network error codes that should trigger a retry
|
|
12
16
|
*/
|
|
@@ -56,6 +60,25 @@ function isRetryableDownloadError(error) {
|
|
|
56
60
|
}
|
|
57
61
|
return false;
|
|
58
62
|
}
|
|
63
|
+
/**
|
|
64
|
+
* Reject `detectImageType()`'s `application/octet-stream` sentinel — the
|
|
65
|
+
* honest "no known image signature (PNG/JPEG/GIF/WebP/BMP/TIFF/SVG/AVIF)
|
|
66
|
+
* matched" fallback (#261/#286). Packaging that sentinel into a data URI
|
|
67
|
+
* hands vision providers (OpenAI/Anthropic/Google) a MIME type they reject
|
|
68
|
+
* outright with an HTTP 400, so every public conversion path that turns
|
|
69
|
+
* unknown bytes into image output must fail loud here first instead of
|
|
70
|
+
* letting the sentinel slip through to a caller.
|
|
71
|
+
*
|
|
72
|
+
* @param mediaType - The MIME type returned by `detectImageType()`.
|
|
73
|
+
* @throws Error if `mediaType` is the octet-stream sentinel.
|
|
74
|
+
*/
|
|
75
|
+
function assertKnownImageType(mediaType) {
|
|
76
|
+
if (mediaType === "application/octet-stream") {
|
|
77
|
+
logger.error("Unable to detect a supported image format from file content");
|
|
78
|
+
throw new Error("Unsupported or corrupted image: no known image signature " +
|
|
79
|
+
"(PNG/JPEG/GIF/WebP/BMP/TIFF/SVG/AVIF) was found in the file content.");
|
|
80
|
+
}
|
|
81
|
+
}
|
|
59
82
|
/**
|
|
60
83
|
* Image processor class for handling provider-specific image formatting
|
|
61
84
|
*/
|
|
@@ -75,7 +98,10 @@ export class ImageProcessor {
|
|
|
75
98
|
throw new Error("Invalid image processing: buffer is empty");
|
|
76
99
|
}
|
|
77
100
|
const mediaType = this.detectImageType(content);
|
|
78
|
-
|
|
101
|
+
// #261/#286 follow-up: fail loud instead of packaging the octet-stream
|
|
102
|
+
// sentinel as a "valid" image (see assertKnownImageType() above).
|
|
103
|
+
assertKnownImageType(mediaType);
|
|
104
|
+
const base64 = ImageProcessor.safeBase64Convert(content, "image processing");
|
|
79
105
|
const dataUri = `data:${mediaType};base64,${base64}`;
|
|
80
106
|
// Validate output before returning
|
|
81
107
|
this.validateProcessOutput(dataUri, base64, mediaType);
|
|
@@ -138,10 +164,15 @@ export class ImageProcessor {
|
|
|
138
164
|
return `data:image/jpeg;base64,${image}`;
|
|
139
165
|
}
|
|
140
166
|
// Handle Buffer - convert to data URI
|
|
141
|
-
const base64 =
|
|
167
|
+
const base64 = ImageProcessor.safeBase64Convert(image, "OpenAI image input");
|
|
142
168
|
return `data:image/jpeg;base64,${base64}`;
|
|
143
169
|
}
|
|
144
170
|
catch (error) {
|
|
171
|
+
// Preserve typed errors (e.g. IMAGE_TOO_LARGE) as-is — don't flatten
|
|
172
|
+
// them into a generic Error and lose the error `code` for callers.
|
|
173
|
+
if (error instanceof NeuroLinkError) {
|
|
174
|
+
throw error;
|
|
175
|
+
}
|
|
145
176
|
logger.error("Failed to process image for OpenAI:", error);
|
|
146
177
|
throw new Error(`Image processing failed for OpenAI: ${error instanceof Error ? error.message : "Unknown error"}`, { cause: error });
|
|
147
178
|
}
|
|
@@ -170,7 +201,7 @@ export class ImageProcessor {
|
|
|
170
201
|
}
|
|
171
202
|
}
|
|
172
203
|
else {
|
|
173
|
-
base64Data =
|
|
204
|
+
base64Data = ImageProcessor.safeBase64Convert(image, "provider image input");
|
|
174
205
|
}
|
|
175
206
|
return {
|
|
176
207
|
mimeType,
|
|
@@ -178,6 +209,11 @@ export class ImageProcessor {
|
|
|
178
209
|
};
|
|
179
210
|
}
|
|
180
211
|
catch (error) {
|
|
212
|
+
// Preserve typed errors (e.g. IMAGE_TOO_LARGE) as-is — don't flatten
|
|
213
|
+
// them into a generic Error and lose the error `code` for callers.
|
|
214
|
+
if (error instanceof NeuroLinkError) {
|
|
215
|
+
throw error;
|
|
216
|
+
}
|
|
181
217
|
logger.error("Failed to process image for Google AI:", error);
|
|
182
218
|
throw new Error(`Image processing failed for Google AI: ${error instanceof Error ? error.message : "Unknown error"}`, { cause: error });
|
|
183
219
|
}
|
|
@@ -206,7 +242,7 @@ export class ImageProcessor {
|
|
|
206
242
|
}
|
|
207
243
|
}
|
|
208
244
|
else {
|
|
209
|
-
base64Data =
|
|
245
|
+
base64Data = ImageProcessor.safeBase64Convert(image, "provider image input");
|
|
210
246
|
}
|
|
211
247
|
return {
|
|
212
248
|
mediaType,
|
|
@@ -214,6 +250,11 @@ export class ImageProcessor {
|
|
|
214
250
|
};
|
|
215
251
|
}
|
|
216
252
|
catch (error) {
|
|
253
|
+
// Preserve typed errors (e.g. IMAGE_TOO_LARGE) as-is — don't flatten
|
|
254
|
+
// them into a generic Error and lose the error `code` for callers.
|
|
255
|
+
if (error instanceof NeuroLinkError) {
|
|
256
|
+
throw error;
|
|
257
|
+
}
|
|
217
258
|
logger.error("Failed to process image for Anthropic:", error);
|
|
218
259
|
throw new Error(`Image processing failed for Anthropic: ${error instanceof Error ? error.message : "Unknown error"}`, { cause: error });
|
|
219
260
|
}
|
|
@@ -238,6 +279,11 @@ export class ImageProcessor {
|
|
|
238
279
|
}
|
|
239
280
|
}
|
|
240
281
|
catch (error) {
|
|
282
|
+
// Preserve typed errors bubbling up from processImageForGoogle/
|
|
283
|
+
// processImageForAnthropic (e.g. IMAGE_TOO_LARGE) as-is.
|
|
284
|
+
if (error instanceof NeuroLinkError) {
|
|
285
|
+
throw error;
|
|
286
|
+
}
|
|
241
287
|
logger.error("Failed to process image for Vertex AI:", error);
|
|
242
288
|
throw new Error(`Image processing failed for Vertex AI: ${error instanceof Error ? error.message : "Unknown error"}`, { cause: error });
|
|
243
289
|
}
|
|
@@ -305,37 +351,111 @@ export class ImageProcessor {
|
|
|
305
351
|
return "image/svg+xml";
|
|
306
352
|
}
|
|
307
353
|
}
|
|
308
|
-
// AVIF:
|
|
354
|
+
// AVIF (ISOBMFF): "ftyp" box at offset 4, major brand at offset 8.
|
|
355
|
+
// Accept the AVIF-spec brand variants — 'avif' (still image), 'avis'
|
|
356
|
+
// (image sequence), and 'avio' (intra-only AV1 image/sequence — the
|
|
357
|
+
// spec lists it under compatible_brands rather than major_brand, but
|
|
358
|
+
// real-world encoders emit it as major_brand too, so it's accepted
|
|
359
|
+
// here) — and compare bytes directly rather than via Buffer.toString()
|
|
360
|
+
// (which would utf-8-decode non-ASCII bytes) so the check is exact
|
|
361
|
+
// (#286).
|
|
309
362
|
if (input.length >= 12) {
|
|
310
|
-
const
|
|
311
|
-
|
|
312
|
-
|
|
363
|
+
const isFtyp = input[4] === 0x66 && // 'f'
|
|
364
|
+
input[5] === 0x74 && // 't'
|
|
365
|
+
input[6] === 0x79 && // 'y'
|
|
366
|
+
input[7] === 0x70; // 'p'
|
|
367
|
+
const brand = String.fromCharCode(input[8], input[9], input[10], input[11]);
|
|
368
|
+
if (isFtyp &&
|
|
369
|
+
(brand === "avif" || brand === "avis" || brand === "avio")) {
|
|
313
370
|
return "image/avif";
|
|
314
371
|
}
|
|
315
372
|
}
|
|
373
|
+
// BMP: "BM" magic (0x42 0x4D)
|
|
374
|
+
if (input[0] === 0x42 && input[1] === 0x4d) {
|
|
375
|
+
return "image/bmp";
|
|
376
|
+
}
|
|
377
|
+
// TIFF: little-endian "II*\0" or big-endian "MM\0*"
|
|
378
|
+
if ((input[0] === 0x49 &&
|
|
379
|
+
input[1] === 0x49 &&
|
|
380
|
+
input[2] === 0x2a &&
|
|
381
|
+
input[3] === 0x00) ||
|
|
382
|
+
(input[0] === 0x4d &&
|
|
383
|
+
input[1] === 0x4d &&
|
|
384
|
+
input[2] === 0x00 &&
|
|
385
|
+
input[3] === 0x2a)) {
|
|
386
|
+
return "image/tiff";
|
|
387
|
+
}
|
|
316
388
|
}
|
|
317
|
-
|
|
389
|
+
// No known image signature matched. Return a safe, honest sentinel
|
|
390
|
+
// rather than mislabeling arbitrary bytes as JPEG (#261): a wrong
|
|
391
|
+
// image/jpeg label silently corrupts the downstream provider request.
|
|
392
|
+
logger.warn("Could not detect image type from magic bytes; using application/octet-stream");
|
|
393
|
+
return "application/octet-stream";
|
|
318
394
|
}
|
|
319
395
|
catch (error) {
|
|
320
396
|
logger.warn("Failed to detect image type, using default:", error);
|
|
321
|
-
return "
|
|
397
|
+
return "application/octet-stream";
|
|
322
398
|
}
|
|
323
399
|
}
|
|
324
400
|
/**
|
|
325
|
-
*
|
|
401
|
+
* Throwing size guard for a raw byte length. Prevents memory exhaustion by
|
|
402
|
+
* rejecting an oversized input BEFORE an unbounded read/allocation happens
|
|
403
|
+
* (#257) — callers that can get a size cheaply (e.g. `fs.stat`) should
|
|
404
|
+
* validate it before ever reading the bytes into memory. Default limit is
|
|
405
|
+
* the canonical `SIZE_LIMITS_BYTES.IMAGE_MAX` (10 MB); there is no public
|
|
406
|
+
* way to raise it — `maxSize` is an internal-only override for advanced
|
|
407
|
+
* callers (e.g. video image inputs use a higher limit).
|
|
408
|
+
*
|
|
409
|
+
* @param size - Byte length to validate
|
|
410
|
+
* @param context - Short label identifying the source (unused in the
|
|
411
|
+
* thrown message today, kept for call-site readability and future use)
|
|
412
|
+
* @param maxSize - Max allowed size in bytes
|
|
413
|
+
* @throws NeuroLinkError (`INVALID_IMAGE_SIZE`) if `size` is not a finite,
|
|
414
|
+
* non-negative number
|
|
415
|
+
* @throws NeuroLinkError (`IMAGE_TOO_LARGE`) if `size` exceeds `maxSize`
|
|
326
416
|
*/
|
|
327
|
-
static
|
|
328
|
-
|
|
329
|
-
|
|
330
|
-
|
|
331
|
-
|
|
332
|
-
|
|
333
|
-
}
|
|
334
|
-
|
|
335
|
-
|
|
336
|
-
|
|
417
|
+
static validateSize(size, context, maxSize = SIZE_LIMITS_BYTES.IMAGE_MAX) {
|
|
418
|
+
// Reject a malformed maxSize before it's ever compared: `size > NaN` and
|
|
419
|
+
// `size > Infinity` (for a negative maxSize) can both evaluate `false`,
|
|
420
|
+
// which would let a corrupted/negative limit bypass the guard entirely.
|
|
421
|
+
if (!Number.isFinite(maxSize) || maxSize < 0) {
|
|
422
|
+
throw ErrorFactory.invalidConfiguration("maxSize", "must be a finite, non-negative byte count", { maxSize, context });
|
|
423
|
+
}
|
|
424
|
+
// Fail closed on malformed sizes first: `NaN > maxSize` and
|
|
425
|
+
// `-1 > maxSize` both evaluate to `false`, which would otherwise let a
|
|
426
|
+
// corrupted stat/header value silently skip the guard below.
|
|
427
|
+
if (!Number.isFinite(size) || size < 0) {
|
|
428
|
+
throw ErrorFactory.invalidImageSize(size);
|
|
429
|
+
}
|
|
430
|
+
if (size > maxSize) {
|
|
431
|
+
const mb = (size / (1024 * 1024)).toFixed(1);
|
|
432
|
+
const maxMb = (maxSize / (1024 * 1024)).toFixed(0);
|
|
433
|
+
throw ErrorFactory.imageTooLarge(mb, maxMb);
|
|
337
434
|
}
|
|
338
435
|
}
|
|
436
|
+
/**
|
|
437
|
+
* Throwing size guard for image buffers. Prevents memory exhaustion by
|
|
438
|
+
* rejecting oversized buffers with a descriptive error BEFORE an unbounded
|
|
439
|
+
* `Buffer.toString("base64")` allocation (#257). Delegates to
|
|
440
|
+
* `validateSize` so buffer-based and pre-read (stat-based) callers share
|
|
441
|
+
* one implementation.
|
|
442
|
+
*
|
|
443
|
+
* @param buffer - Image buffer to validate
|
|
444
|
+
* @param context - Short label identifying the source (see `validateSize`)
|
|
445
|
+
* @param maxSize - Max allowed size in bytes
|
|
446
|
+
* @throws NeuroLinkError (`IMAGE_TOO_LARGE`) if the buffer exceeds `maxSize`
|
|
447
|
+
*/
|
|
448
|
+
static validateBufferSize(buffer, context, maxSize = SIZE_LIMITS_BYTES.IMAGE_MAX) {
|
|
449
|
+
ImageProcessor.validateSize(buffer.length, context, maxSize);
|
|
450
|
+
}
|
|
451
|
+
/**
|
|
452
|
+
* Size-guarded Buffer → base64. Use in place of `buffer.toString("base64")`
|
|
453
|
+
* on any path that converts a caller-supplied image buffer (#257).
|
|
454
|
+
*/
|
|
455
|
+
static safeBase64Convert(buffer, context, maxSize = SIZE_LIMITS_BYTES.IMAGE_MAX) {
|
|
456
|
+
ImageProcessor.validateBufferSize(buffer, context, maxSize);
|
|
457
|
+
return buffer.toString("base64");
|
|
458
|
+
}
|
|
339
459
|
/**
|
|
340
460
|
* Validate image format
|
|
341
461
|
*/
|
|
@@ -349,6 +469,12 @@ export class ImageProcessor {
|
|
|
349
469
|
"image/tiff",
|
|
350
470
|
"image/svg+xml",
|
|
351
471
|
"image/avif",
|
|
472
|
+
// Deliberately excludes "application/octet-stream": that's
|
|
473
|
+
// detectImageType()'s honest sentinel for bytes matching no known
|
|
474
|
+
// image signature (#261), not a real image format. Vision providers
|
|
475
|
+
// (OpenAI/Anthropic/Google) reject it outright with an HTTP 400, so
|
|
476
|
+
// process() must fail loud instead of packaging it as a valid image
|
|
477
|
+
// (see the octet-stream guard in process() below).
|
|
352
478
|
];
|
|
353
479
|
return supportedFormats.includes(mediaType.toLowerCase());
|
|
354
480
|
}
|
|
@@ -429,7 +555,19 @@ export class ImageProcessor {
|
|
|
429
555
|
*/
|
|
430
556
|
static processImage(image, provider, model) {
|
|
431
557
|
try {
|
|
558
|
+
// #257: reject an oversized buffer before any format-detection or
|
|
559
|
+
// conversion work runs on it — mirrors the guard every
|
|
560
|
+
// `processImageForX()` branch below already applies via
|
|
561
|
+
// `safeBase64Convert()`, so the typed IMAGE_TOO_LARGE error surfaces
|
|
562
|
+
// here too instead of being pre-empted by the octet-stream check next.
|
|
563
|
+
if (Buffer.isBuffer(image)) {
|
|
564
|
+
ImageProcessor.validateBufferSize(image, "image processing");
|
|
565
|
+
}
|
|
432
566
|
const mediaType = ImageProcessor.detectImageType(image);
|
|
567
|
+
// #261/#286 follow-up: reject the octet-stream sentinel here too — the
|
|
568
|
+
// returned ProcessedImage.mediaType is what callers use to build a
|
|
569
|
+
// data URI, so this is a separate public path that must not emit it.
|
|
570
|
+
assertKnownImageType(mediaType);
|
|
433
571
|
const size = typeof image === "string"
|
|
434
572
|
? Buffer.byteLength(image, "base64")
|
|
435
573
|
: image.length;
|
|
@@ -467,7 +605,7 @@ export class ImageProcessor {
|
|
|
467
605
|
: image;
|
|
468
606
|
}
|
|
469
607
|
else {
|
|
470
|
-
data =
|
|
608
|
+
data = ImageProcessor.safeBase64Convert(image, "image input");
|
|
471
609
|
}
|
|
472
610
|
format = "base64";
|
|
473
611
|
}
|
|
@@ -479,6 +617,11 @@ export class ImageProcessor {
|
|
|
479
617
|
};
|
|
480
618
|
}
|
|
481
619
|
catch (error) {
|
|
620
|
+
// Preserve typed errors (e.g. IMAGE_TOO_LARGE) bubbling up from the
|
|
621
|
+
// provider-specific branches above as-is.
|
|
622
|
+
if (error instanceof NeuroLinkError) {
|
|
623
|
+
throw error;
|
|
624
|
+
}
|
|
482
625
|
logger.error(`Failed to process image for ${provider}:`, error);
|
|
483
626
|
throw new Error(`Image processing failed: ${error instanceof Error ? error.message : "Unknown error"}`, { cause: error });
|
|
484
627
|
}
|
|
@@ -597,10 +740,16 @@ export const imageUtils = {
|
|
|
597
740
|
return parseFloat((bytes / Math.pow(k, i)).toFixed(2)) + " " + sizes[i];
|
|
598
741
|
},
|
|
599
742
|
/**
|
|
600
|
-
* Convert Buffer to base64 string
|
|
743
|
+
* Convert Buffer to base64 string. Guarded by the same #257 size limit as
|
|
744
|
+
* every other conversion site (default `SIZE_LIMITS_BYTES.IMAGE_MAX`);
|
|
745
|
+
* pass `maxBytes` to override for a caller with a legitimate need for a
|
|
746
|
+
* different ceiling rather than silently accepting an unbounded buffer.
|
|
747
|
+
* `imageUtils`/`ImageProcessor` are internal-only — not re-exported from
|
|
748
|
+
* `src/lib/index.ts` or the package's public `exports` map — so this
|
|
749
|
+
* default does not change behavior for any external SDK consumer.
|
|
601
750
|
*/
|
|
602
|
-
bufferToBase64: (buffer) => {
|
|
603
|
-
return
|
|
751
|
+
bufferToBase64: (buffer, maxBytes = SIZE_LIMITS_BYTES.IMAGE_MAX) => {
|
|
752
|
+
return ImageProcessor.safeBase64Convert(buffer, "buffer-to-base64", maxBytes);
|
|
604
753
|
},
|
|
605
754
|
/**
|
|
606
755
|
* Convert base64 string to Buffer
|
|
@@ -610,6 +759,15 @@ export const imageUtils = {
|
|
|
610
759
|
const cleanBase64 = base64.includes(",") ? base64.split(",")[1] : base64;
|
|
611
760
|
return Buffer.from(cleanBase64, "base64");
|
|
612
761
|
},
|
|
762
|
+
/**
|
|
763
|
+
* Redact `filePath` (both as given and its `path.resolve()`'d form) from
|
|
764
|
+
* an error message. Exposed directly (not just used internally by
|
|
765
|
+
* {@link fileToBase64DataUri}) so the resolved-path branch can be verified
|
|
766
|
+
* by a deterministic unit test — real `fs` errors only ever embed the
|
|
767
|
+
* literal path as passed, never a resolved variant, so that branch isn't
|
|
768
|
+
* otherwise observable through the async file-reading API.
|
|
769
|
+
*/
|
|
770
|
+
redactPathFromMessage,
|
|
613
771
|
/**
|
|
614
772
|
* Convert file path to base64 data URI
|
|
615
773
|
*/
|
|
@@ -629,11 +787,38 @@ export const imageUtils = {
|
|
|
629
787
|
// Enhanced MIME detection: try buffer content first, fallback to filename
|
|
630
788
|
const mimeType = ImageProcessor.detectImageType(buffer) ||
|
|
631
789
|
ImageProcessor.detectImageType(filePath);
|
|
790
|
+
// #261/#286 follow-up: reject the octet-stream sentinel here too —
|
|
791
|
+
// this is a third public path (alongside process() and
|
|
792
|
+
// processImage()) that can otherwise package unknown bytes into a
|
|
793
|
+
// `data:application/octet-stream;base64,...` URI a vision provider
|
|
794
|
+
// rejects outright.
|
|
795
|
+
assertKnownImageType(mimeType);
|
|
632
796
|
const base64 = buffer.toString("base64");
|
|
633
797
|
return `data:${mimeType};base64,${base64}`;
|
|
634
798
|
}
|
|
635
799
|
catch (error) {
|
|
636
|
-
|
|
800
|
+
// Full path stays in the debug log; the thrown (potentially
|
|
801
|
+
// client-facing) error only gets the basename. The underlying fs
|
|
802
|
+
// error's own message is scrubbed too — Node's ENOENT/EACCES text
|
|
803
|
+
// embeds the full path verbatim (e.g. "no such file or directory,
|
|
804
|
+
// stat '/full/path'"), which would otherwise leak the host's
|
|
805
|
+
// directory layout right back through the "reason" suffix. The raw
|
|
806
|
+
// `error` is never attached as `cause` either — that would leave the
|
|
807
|
+
// unredacted path reachable via `error.cause.message` for anything
|
|
808
|
+
// that walks the cause chain (cause-aware logging, telemetry); a
|
|
809
|
+
// sanitized copy (via the shared `sanitizeErrorCause`) is attached
|
|
810
|
+
// instead, mirroring `urlToBase64DataUri` below.
|
|
811
|
+
logger.debug("fileToBase64DataUri failed", { filePath, error });
|
|
812
|
+
const safeName = basename(filePath);
|
|
813
|
+
const cause = sanitizeErrorCause(error, { filePath });
|
|
814
|
+
const reason = error instanceof Error ? cause.message : "Unknown error";
|
|
815
|
+
// Assigned to a variable before throwing (rather than `throw new
|
|
816
|
+
// Error(...)` inline) so the sanitized `cause` — a deliberately NEW,
|
|
817
|
+
// redacted Error — isn't mistaken by tooling for a dropped/altered
|
|
818
|
+
// reference to the originally caught `error`; the raw error is never
|
|
819
|
+
// reachable from the thrown result at all, by design.
|
|
820
|
+
const wrapped = new Error(`Failed to convert file to base64 (${safeName}): ${reason}`, { cause });
|
|
821
|
+
throw wrapped;
|
|
637
822
|
}
|
|
638
823
|
},
|
|
639
824
|
/**
|
|
@@ -719,14 +904,31 @@ export const imageUtils = {
|
|
|
719
904
|
maxDelay: SYSTEM_LIMITS.DEFAULT_MAX_DELAY,
|
|
720
905
|
retryCondition: isRetryableDownloadError,
|
|
721
906
|
onRetry: (attempt, error) => {
|
|
722
|
-
|
|
907
|
+
// Some fetch/DNS/TLS errors embed the full request URL (including
|
|
908
|
+
// a presigned query token) in `error.message` — redact that text
|
|
909
|
+
// too, not just the explicit `url` interpolation below. The
|
|
910
|
+
// non-Error branch is sanitized too: String(error) on a thrown
|
|
911
|
+
// non-Error value (e.g. a raw string or object) can just as
|
|
912
|
+
// easily carry the same signed URL.
|
|
913
|
+
const message = redactUrlsInText(error instanceof Error ? error.message : String(error));
|
|
723
914
|
const attemptsLeft = maxAttempts - attempt;
|
|
724
|
-
logger.warn(`⚠️ Image download attempt ${attempt} failed for ${url}: ${message}. ${attemptsLeft} ${attemptsLeft === 1 ? "attempt" : "attempts"} remaining...`);
|
|
915
|
+
logger.warn(`⚠️ Image download attempt ${attempt} failed for ${redactUrlForError(url)}: ${message}. ${attemptsLeft} ${attemptsLeft === 1 ? "attempt" : "attempts"} remaining...`);
|
|
725
916
|
},
|
|
726
917
|
});
|
|
727
918
|
}
|
|
728
919
|
catch (error) {
|
|
729
|
-
|
|
920
|
+
// Query string / fragment stripped — a presigned S3/GCS URL or SAS
|
|
921
|
+
// token embedded there must not be echoed into a thrown error. The
|
|
922
|
+
// underlying error's own message is scrubbed too, since fetch/DNS/TLS
|
|
923
|
+
// errors often embed the full URL (query string and all) themselves.
|
|
924
|
+
// The raw `error` is never attached as `cause` — that would leave the
|
|
925
|
+
// unredacted URL reachable via `error.cause.message` for anything
|
|
926
|
+
// that walks the cause chain (cause-aware logging, telemetry); a
|
|
927
|
+
// sanitized copy is attached instead.
|
|
928
|
+
const sourceUrl = redactUrlForError(url);
|
|
929
|
+
const cause = sanitizeErrorCause(error);
|
|
930
|
+
const wrapped = new Error(`Failed to download and convert URL to base64 (${sourceUrl}): ${cause.message}`, { cause });
|
|
931
|
+
throw wrapped;
|
|
730
932
|
}
|
|
731
933
|
},
|
|
732
934
|
/**
|
|
@@ -61,11 +61,110 @@ export declare function safeDebugSerialize(value: unknown, maxLen?: number): str
|
|
|
61
61
|
* for diagnostics. Use this for any `baseURL`/endpoint a caller may have
|
|
62
62
|
* supplied with inline credentials. The match is global, so every `//…@`
|
|
63
63
|
* authority in the string is redacted (e.g. a proxy chain or a URL embedded
|
|
64
|
-
* in a query parameter), not just the first.
|
|
64
|
+
* in a query parameter), not just the first. IPv6 literal hosts (`//u:p@[::1]/x`)
|
|
65
|
+
* and percent-encoded userinfo (`//u:p%40x@host/x`) are handled naturally —
|
|
66
|
+
* neither `[`/`]` nor `%` stop either pass below, so the bracketed host or
|
|
67
|
+
* encoded byte just rides along as part of the (non-redacted) remainder.
|
|
68
|
+
*
|
|
69
|
+
* This is best-effort defense-in-depth for free-form logging, deliberately
|
|
70
|
+
* regex-based so it can scrub *multiple* embedded authorities out of
|
|
71
|
+
* arbitrary text (a real URL parser only ever parses one URL at a time).
|
|
72
|
+
* It is NOT the primary protection for user-facing errors — that's
|
|
73
|
+
* {@link redactUrlForError}, which parses the single input via `new URL()`
|
|
74
|
+
* and reconstructs from protocol+host+pathname, never touching userinfo at
|
|
75
|
+
* all. Prefer that function whenever the input is known to be exactly one
|
|
76
|
+
* URL.
|
|
77
|
+
*
|
|
78
|
+
* Two passes, deliberately:
|
|
79
|
+
* 1. Well-formed authorities — credentials with no raw `/` — bounded at the
|
|
80
|
+
* authority's first `/` (or string end). Greedy backtracking naturally
|
|
81
|
+
* absorbs multiple `@` segments within one authority (e.g. `//a@b@host`
|
|
82
|
+
* → `//***@host`) while leaving a second, separate `//...@` authority
|
|
83
|
+
* elsewhere in the string (e.g. one embedded in a query parameter)
|
|
84
|
+
* untouched, because pass 1's own exclusion of `/` stops it at the first
|
|
85
|
+
* `/`, before it can ever reach a nested authority.
|
|
86
|
+
* 2. Malformed authorities whose credentials contain an unescaped `/`
|
|
87
|
+
* (e.g. `//user:sec/ret@host/path`) — pass 1 can't find an `@` before
|
|
88
|
+
* hitting that `/`, so this pass allows `/` and redacts up to the last
|
|
89
|
+
* remaining `@`. It stops before any *nested* `//` (via a negative
|
|
90
|
+
* lookahead) rather than excluding a specific character like `*` —
|
|
91
|
+
* excluding a literal character is what let a password containing both
|
|
92
|
+
* `/` and `*` (e.g. `//user:pa*ss/word@host/path`, legal under RFC 3986)
|
|
93
|
+
* slip through both passes entirely in an earlier version of this
|
|
94
|
+
* function.
|
|
65
95
|
*
|
|
66
96
|
* @param url - The URL (or URL-shaped string) to redact.
|
|
67
97
|
*/
|
|
68
98
|
export declare function redactUrlCredentials(url: string): string;
|
|
99
|
+
/**
|
|
100
|
+
* Reduce a URL to its origin + pathname for safe inclusion in error messages
|
|
101
|
+
* and logs, dropping the query string and fragment — where presigned-URL
|
|
102
|
+
* signatures, SAS tokens, and other secrets commonly live — while keeping
|
|
103
|
+
* enough of the URL (scheme, host, path) for the error to stay diagnostic.
|
|
104
|
+
*
|
|
105
|
+
* Falls back to a best-effort redacted string (query/fragment and
|
|
106
|
+
* `user:pass@` stripped) when the input isn't a parseable absolute URL.
|
|
107
|
+
* Either way the result is capped at 200 chars so a very long host/path
|
|
108
|
+
* can't blow up the error message.
|
|
109
|
+
*
|
|
110
|
+
* @param url - The URL (or URL-shaped string) to redact.
|
|
111
|
+
*/
|
|
112
|
+
export declare function redactUrlForError(url: string): string;
|
|
113
|
+
/**
|
|
114
|
+
* Scrub any embedded absolute URLs out of free-form text — typically the
|
|
115
|
+
* `.message` of an underlying network/DNS/TLS error thrown by `fetch`,
|
|
116
|
+
* undici's `request()`, or Node's resolver, all of which frequently embed
|
|
117
|
+
* the full request URL (query string, credentials, and all) verbatim.
|
|
118
|
+
* Interpolating that message raw into a thrown or logged error would leak a
|
|
119
|
+
* presigned URL's token even after the explicit URL argument was already
|
|
120
|
+
* redacted via {@link redactUrlForError}. Finds each `scheme://…` run in the
|
|
121
|
+
* text and replaces it with its redacted form.
|
|
122
|
+
*
|
|
123
|
+
* @param text - Free-form text (typically `Error.message`) to scrub.
|
|
124
|
+
*/
|
|
125
|
+
export declare function redactUrlsInText(text: string): string;
|
|
126
|
+
/**
|
|
127
|
+
* Redact every occurrence of `filePath` from an error message, checking both
|
|
128
|
+
* the path exactly as given and its `path.resolve()`'d form.
|
|
129
|
+
*
|
|
130
|
+
* Node's own fs errors (ENOENT/EACCES) embed the exact string passed to the
|
|
131
|
+
* fs call, so a plain substring match usually suffices — but wrapped or
|
|
132
|
+
* lower-level errors sometimes normalize the path first (relative→absolute,
|
|
133
|
+
* trailing-slash variations), which an exact-match-only redaction would miss.
|
|
134
|
+
* This is defense-in-depth, not a guarantee: a `realpath`-resolved (symlink-
|
|
135
|
+
* following) variant can still differ from both forms and would slip through.
|
|
136
|
+
*
|
|
137
|
+
* @param message - The error message to scrub.
|
|
138
|
+
* @param filePath - The path to redact (both as given and resolved).
|
|
139
|
+
*/
|
|
140
|
+
export declare function redactPathFromMessage(message: string, filePath: string): string;
|
|
141
|
+
/**
|
|
142
|
+
* Build a sanitized copy of a network/URL/file error, safe to attach as
|
|
143
|
+
* `Error.cause` on a rethrow.
|
|
144
|
+
*
|
|
145
|
+
* Redacting only the OUTER thrown error's message isn't enough: if the RAW
|
|
146
|
+
* underlying error is attached as `cause`, anything that walks the cause
|
|
147
|
+
* chain (cause-aware logging, telemetry spans) can still recover an
|
|
148
|
+
* unredacted URL/path/secret via `error.cause.message`. This returns a NEW
|
|
149
|
+
* `Error` whose message has been through {@link redactUrlsInText} — and,
|
|
150
|
+
* when `options.filePath` is supplied, {@link redactPathFromMessage} too, so
|
|
151
|
+
* a filesystem error's ENOENT/EACCES text doesn't leak the host's directory
|
|
152
|
+
* layout via the cause chain the way the outer message already avoids.
|
|
153
|
+
* `.name` and `.code` (`NodeJS.ErrnoException`) are copied over so retry
|
|
154
|
+
* classification (e.g. `isRetryableNetworkError`) still works when it
|
|
155
|
+
* inspects the cause. The raw original is never referenced by the result.
|
|
156
|
+
*
|
|
157
|
+
* Handles non-`Error` thrown values too (a raw string/object can just as
|
|
158
|
+
* easily carry an unredacted URL or path) by running their `String()` form
|
|
159
|
+
* through the same redaction before wrapping.
|
|
160
|
+
*
|
|
161
|
+
* @param error - The raw underlying error (or thrown value) to sanitize.
|
|
162
|
+
* @param options - `filePath`: a known filesystem path to redact to its
|
|
163
|
+
* basename, in addition to URL redaction.
|
|
164
|
+
*/
|
|
165
|
+
export declare function sanitizeErrorCause(error: unknown, options?: {
|
|
166
|
+
filePath?: string;
|
|
167
|
+
}): Error;
|
|
69
168
|
/**
|
|
70
169
|
* Recursively sanitize a record/array, returning a structurally identical
|
|
71
170
|
* value with sensitive keys redacted and string values run through
|