@juspay/neurolink 9.94.3 → 9.94.5

package/dist/utils/logSanitize.js

@@ -15,6 +15,7 @@
  *     jina_ (Jina), fish- (Fish Audio)
  *   - Generic key=value pairs: api_key=…, access_token: …, secret_key=…
  */
+import { basename, resolve as resolvePath } from "path";
 const TOKEN_PREFIXES = [
     "sk",
     "pk",
@@ -156,12 +157,161 @@ export function safeDebugSerialize(value, maxLen = 10_000) {
  * for diagnostics. Use this for any `baseURL`/endpoint a caller may have
  * supplied with inline credentials. The match is global, so every `//…@`
  * authority in the string is redacted (e.g. a proxy chain or a URL embedded
- * in a query parameter), not just the first.
+ * in a query parameter), not just the first. IPv6 literal hosts (`//u:p@[::1]/x`)
+ * and percent-encoded userinfo (`//u:p%40x@host/x`) are handled naturally —
+ * neither `[`/`]` nor `%` stop either pass below, so the bracketed host or
+ * encoded byte just rides along as part of the (non-redacted) remainder.
+ *
+ * This is best-effort defense-in-depth for free-form logging, deliberately
+ * regex-based so it can scrub *multiple* embedded authorities out of
+ * arbitrary text (a real URL parser only ever parses one URL at a time).
+ * It is NOT the primary protection for user-facing errors — that's
+ * {@link redactUrlForError}, which parses the single input via `new URL()`
+ * and reconstructs from protocol+host+pathname, never touching userinfo at
+ * all. Prefer that function whenever the input is known to be exactly one
+ * URL.
+ *
+ * Two passes, deliberately:
+ *  1. Well-formed authorities — credentials with no raw `/` — bounded at the
+ *     authority's first `/` (or string end). Greedy backtracking naturally
+ *     absorbs multiple `@` segments within one authority (e.g. `//a@b@host`
+ *     → `//***@host`) while leaving a second, separate `//...@` authority
+ *     elsewhere in the string (e.g. one embedded in a query parameter)
+ *     untouched, because pass 1's own exclusion of `/` stops it at the first
+ *     `/`, before it can ever reach a nested authority.
+ *  2. Malformed authorities whose credentials contain an unescaped `/`
+ *     (e.g. `//user:sec/ret@host/path`) — pass 1 can't find an `@` before
+ *     hitting that `/`, so this pass allows `/` and redacts up to the last
+ *     remaining `@`. It stops before any *nested* `//` (via a negative
+ *     lookahead) rather than excluding a specific character like `*` —
+ *     excluding a literal character is what let a password containing both
+ *     `/` and `*` (e.g. `//user:pa*ss/word@host/path`, legal under RFC 3986)
+ *     slip through both passes entirely in an earlier version of this
+ *     function.
  *
  * @param url - The URL (or URL-shaped string) to redact.
  */
 export function redactUrlCredentials(url) {
-    return url.replace(/\/\/[^/@]+@/g, "//***@");
+    const wellFormed = url.replace(/\/\/[^/\s"'<>]*@/g, "//***@");
+    return wellFormed.replace(/\/\/(?:(?!\/\/)[^\s"'<>])*@/g, "//***@");
+}
+/**
+ * Reduce a URL to its origin + pathname for safe inclusion in error messages
+ * and logs, dropping the query string and fragment — where presigned-URL
+ * signatures, SAS tokens, and other secrets commonly live — while keeping
+ * enough of the URL (scheme, host, path) for the error to stay diagnostic.
+ *
+ * Falls back to a best-effort redacted string (query/fragment and
+ * `user:pass@` stripped) when the input isn't a parseable absolute URL.
+ * Either way the result is capped at 200 chars so a very long host/path
+ * can't blow up the error message.
+ *
+ * @param url - The URL (or URL-shaped string) to redact.
+ */
+export function redactUrlForError(url) {
+    const truncate = (s) => s.length > 200 ? `${s.slice(0, 200)}…` : s;
+    try {
+        const parsed = new URL(url);
+        // Reconstruct explicitly from protocol + host + pathname — never
+        // parsed.username/parsed.password, and never parsed.href (which
+        // re-serializes any embedded userinfo) — so a `user:pass@host` in the
+        // input can't leak into a diagnostic error message.
+        return truncate(`${parsed.protocol}//${parsed.host}${parsed.pathname}`);
+    }
+    catch {
+        // Unparseable input: `new URL()` gave us nothing to work with, so
+        // best-effort drop the query/fragment (where secrets like tokens live)
+        // and strip any `//user:pass@` segment before falling back to the raw
+        // string — never echo the input completely unredacted.
+        const withoutQueryOrFragment = url.split(/[?#]/)[0];
+        return truncate(redactUrlCredentials(withoutQueryOrFragment));
+    }
+}
+/**
+ * Scrub any embedded absolute URLs out of free-form text — typically the
+ * `.message` of an underlying network/DNS/TLS error thrown by `fetch`,
+ * undici's `request()`, or Node's resolver, all of which frequently embed
+ * the full request URL (query string, credentials, and all) verbatim.
+ * Interpolating that message raw into a thrown or logged error would leak a
+ * presigned URL's token even after the explicit URL argument was already
+ * redacted via {@link redactUrlForError}. Finds each `scheme://…` run in the
+ * text and replaces it with its redacted form.
+ *
+ * @param text - Free-form text (typically `Error.message`) to scrub.
+ */
+export function redactUrlsInText(text) {
+    return text.replace(/[a-zA-Z][a-zA-Z0-9+.-]*:\/\/[^\s"'<>)]+/g, (match) => redactUrlForError(match));
+}
+/**
+ * Redact every occurrence of `filePath` from an error message, checking both
+ * the path exactly as given and its `path.resolve()`'d form.
+ *
+ * Node's own fs errors (ENOENT/EACCES) embed the exact string passed to the
+ * fs call, so a plain substring match usually suffices — but wrapped or
+ * lower-level errors sometimes normalize the path first (relative→absolute,
+ * trailing-slash variations), which an exact-match-only redaction would miss.
+ * This is defense-in-depth, not a guarantee: a `realpath`-resolved (symlink-
+ * following) variant can still differ from both forms and would slip through.
+ *
+ * @param message  - The error message to scrub.
+ * @param filePath - The path to redact (both as given and resolved).
+ */
+export function redactPathFromMessage(message, filePath) {
+    const safeName = basename(filePath);
+    // Longest-first: the resolved (absolute) form always contains the relative
+    // form as a substring, so redacting the relative form first would only
+    // strip the trailing segment of an absolute path and leave its parent
+    // directory behind (e.g. "/Users/me/project/foo/bar.png" redacting
+    // "foo/bar.png" first leaves "/Users/me/project/bar.png"). Redacting the
+    // longer resolved variant first removes the whole absolute path in one
+    // pass, so the shorter relative pass below has nothing left to match.
+    return message
+        .split(resolvePath(filePath))
+        .join(safeName)
+        .split(filePath)
+        .join(safeName);
+}
+/**
+ * Build a sanitized copy of a network/URL/file error, safe to attach as
+ * `Error.cause` on a rethrow.
+ *
+ * Redacting only the OUTER thrown error's message isn't enough: if the RAW
+ * underlying error is attached as `cause`, anything that walks the cause
+ * chain (cause-aware logging, telemetry spans) can still recover an
+ * unredacted URL/path/secret via `error.cause.message`. This returns a NEW
+ * `Error` whose message has been through {@link redactUrlsInText} — and,
+ * when `options.filePath` is supplied, {@link redactPathFromMessage} too, so
+ * a filesystem error's ENOENT/EACCES text doesn't leak the host's directory
+ * layout via the cause chain the way the outer message already avoids.
+ * `.name` and `.code` (`NodeJS.ErrnoException`) are copied over so retry
+ * classification (e.g. `isRetryableNetworkError`) still works when it
+ * inspects the cause. The raw original is never referenced by the result.
+ *
+ * Handles non-`Error` thrown values too (a raw string/object can just as
+ * easily carry an unredacted URL or path) by running their `String()` form
+ * through the same redaction before wrapping.
+ *
+ * @param error   - The raw underlying error (or thrown value) to sanitize.
+ * @param options - `filePath`: a known filesystem path to redact to its
+ *                  basename, in addition to URL redaction.
+ */
+export function sanitizeErrorCause(error, options) {
+    const redact = (text) => {
+        const urlSafe = redactUrlsInText(text);
+        return options?.filePath
+            ? redactPathFromMessage(urlSafe, options.filePath)
+            : urlSafe;
+    };
+    if (!(error instanceof Error)) {
+        return new Error(redact(String(error)));
+    }
+    const sanitized = new Error(redact(error.message));
+    sanitized.name = error.name;
+    const code = error.code;
+    if (code !== undefined) {
+        sanitized.code = code;
+    }
+    return sanitized;
 }
 /**
  * Recursively sanitize a record/array, returning a structurally identical

package/dist/utils/messageBuilder.js

@@ -1,4 +1,5 @@
 import { existsSync, readFileSync, statSync } from "fs";
+import { readFile as readFileAsync, stat as statAsync } from "fs/promises";
 import { getGlobalDispatcher, interceptors, request } from "undici";
 import { MultimodalLogger, ProviderImageAdapter, } from "../adapters/providerImageAdapter.js";
 import { CONVERSATION_INSTRUCTIONS, STRUCTURED_OUTPUT_INSTRUCTIONS, } from "../config/conversationMemory.js";
@@ -6,8 +7,10 @@ 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 { tracers, ATTR, withSpan } from "../telemetry/index.js";
+import { NeuroLinkError } from "./errorHandling.js";
 import { FileDetector } from "./fileDetector.js";
 import { getImageCache } from "./imageCache.js";
+import { ImageProcessor } from "./imageProcessor.js";
 import { logger } from "./logger.js";
 import { PDFImageConverter, PDFProcessor } from "./pdfProcessor.js";
 import { urlDownloadRateLimiter } from "./rateLimiter.js";
@@ -1396,13 +1399,38 @@ function detectMimeTypeFromBuffer(buffer) {
 /**
  * Convert file path to raw base64 string.
  * Returns raw base64 (not a data: URI) to avoid SSRF validation in AI SDK v6.
+ * Uses async fs so a large/slow-filesystem image never blocks the event loop
+ * while the size guard (below) or the read itself is in flight.
  */
-function convertFilePathToBase64(filePath) {
-    if (!existsSync(filePath)) {
-        throw new Error(`Image file not found: ${filePath}`);
+async function convertFilePathToBase64(filePath) {
+    let stats;
+    try {
+        stats = await statAsync(filePath);
     }
-    const buffer = readFileSync(filePath);
-    return buffer.toString("base64");
+    catch (error) {
+        // Only ENOENT means "not found" — EACCES/ELOOP/etc. are real access or
+        // filesystem problems that must not be misreported as a missing file.
+        const code = error instanceof Error
+            ? error.code
+            : undefined;
+        if (code === "ENOENT") {
+            throw new Error(`Image file not found: ${filePath}`, { cause: error });
+        }
+        throw new Error(`Cannot access image file ${filePath}: ${error instanceof Error ? error.message : String(error)}`, { cause: error });
+    }
+    if (!stats.isFile()) {
+        throw new Error(`Image path is not a file: ${filePath}`);
+    }
+    // #257: check the file size via stat BEFORE reading it into memory, so a
+    // huge image path never triggers an unbounded read + base64 allocation.
+    // Delegates to ImageProcessor's shared guard (no local reimplementation).
+    const context = `image file "${filePath}"`;
+    ImageProcessor.validateSize(stats.size, context);
+    const buffer = await readFileAsync(filePath);
+    // TOCTOU: the file can grow, or a symlink can be swapped, between the stat
+    // above and this read completing. Re-validate the bytes actually read
+    // (not just the pre-read stat) before the base64 allocation.
+    return ImageProcessor.safeBase64Convert(buffer, context);
 }
 /**
  * Process a single image input and convert to raw base64 format.
@@ -1413,7 +1441,7 @@ function convertFilePathToBase64(filePath) {
  * Passing raw base64 avoids this because `new URL(base64string)` throws and
  * the SDK treats the string as inline base64 data instead.
  */
-function processImageToBase64(image, index) {
+async function processImageToBase64(image, index) {
     let imageData;
     let mimeType = "image/jpeg"; // Default mime type
     if (typeof image === "string") {
@@ -1444,7 +1472,7 @@ function processImageToBase64(image, index) {
         else {
             // File path string - convert to raw base64
             try {
-                imageData = convertFilePathToBase64(image);
+                imageData = await convertFilePathToBase64(image);
                 mimeType = getMimeTypeFromExtension(image);
             }
             catch (error) {
@@ -1452,6 +1480,11 @@ function processImageToBase64(image, index) {
                     index,
                     filePath: image,
                 });
+                // Preserve typed errors (e.g. IMAGE_TOO_LARGE) as-is — don't flatten
+                // them into a generic Error and lose the error `code` for callers.
+                if (error instanceof NeuroLinkError) {
+                    throw error;
+                }
                 throw new Error(`Failed to convert file path to base64: ${image}. ${error}`, { cause: error });
             }
         }
@@ -1462,6 +1495,9 @@ function processImageToBase64(image, index) {
         if (detectedMimeType) {
             mimeType = detectedMimeType;
         }
+        // #257: guard the buffer size before the unbounded base64 conversion.
+        // Delegates to ImageProcessor's shared guard (no local reimplementation).
+        ImageProcessor.validateBufferSize(image, `image input at index ${index}`);
         imageData = image.toString("base64");
     }
     return { imageData, mimeType };
@@ -1524,11 +1560,13 @@ async function convertSimpleImagesToProviderFormat(text, images, provider, _mode
     const content = [
         { type: "text", text: enhancedText },
     ];
-    // Process all images (including downloaded URLs) for Vercel AI SDK
-    actualImages.forEach(({ data: image }, index) => {
+    // Process all images (including downloaded URLs) for Vercel AI SDK.
+    // Sequential for...of (not Promise.all) to preserve image ordering and
+    // keep the original forEach's one-at-a-time error semantics.
+    for (const [index, { data: image }] of actualImages.entries()) {
         try {
             // Use helper function to process image and reduce nesting depth
-            const { imageData, mimeType } = processImageToBase64(image, index);
+            const { imageData, mimeType } = await processImageToBase64(image, index);
             content.push({
                 type: "image",
                 image: imageData,
@@ -1542,7 +1580,7 @@ async function convertSimpleImagesToProviderFormat(text, images, provider, _mode
             });
             throw error;
         }
-    });
+    }
     return content;
 }
 /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "9.94.3",
+  "version": "9.94.5",
   "packageManager": "pnpm@10.15.1",
   "description": "TypeScript AI SDK with 24+ LLM providers behind one consistent API. MCP-native (connect any MCP server), voice TTS/STT/realtime, RAG, agents, memory, context compaction. OpenAI · Anthropic · Gemini · Bedrock · Azure · Ollama · DeepSeek · NVIDIA NIM and more.",
   "author": {