refactor(scripts): extract sanitizePromptInput into a shared module

scripts/bedrock_classifier.ts

@@ -9,6 +9,7 @@ import {
 } from "@aws-sdk/client-bedrock-runtime";
 import { ClassificationResult, LabelTaxonomy } from "./data_models.js";
 import { retryWithBackoff } from "./retry_utils.js";
+import { sanitizePromptInput } from "./sanitize_utils.js";
 
 const MODEL_ID = "us.anthropic.claude-sonnet-4-20250514-v1:0";
 const MAX_TOKENS = 2048;
@@ -19,49 +20,6 @@ const TOP_P = 0.9;
 const MAX_TITLE_LENGTH = 500;
 const MAX_BODY_LENGTH = 10000;
 
-/**
- * Sanitize user input to prevent prompt injection attacks
- */
-function sanitizePromptInput(input: string, maxLength: number): string {
-  if (!input) {
-    return "";
-  }
-
-  // Truncate to maximum length
-  let sanitized = input.substring(0, maxLength);
-
-  // Remove potential prompt injection patterns
-  const dangerousPatterns = [
-    /ignore\s+(all\s+)?(previous|above|prior)\s+instructions?/gi,
-    /disregard\s+(all\s+)?(previous|above|prior)\s+instructions?/gi,
-    /forget\s+(all\s+)?(previous|above|prior)\s+instructions?/gi,
-    /new\s+instructions?:/gi,
-    /system\s*:/gi,
-    /assistant\s*:/gi,
-    /\[SYSTEM\]/gi,
-    /\[ASSISTANT\]/gi,
-    /\<\|im_start\|\>/gi,
-    /\<\|im_end\|\>/gi,
-  ];
-
-  for (const pattern of dangerousPatterns) {
-    sanitized = sanitized.replace(pattern, "[REDACTED]");
-  }
-
-  // Escape backticks that could break JSON formatting
-  sanitized = sanitized.replace(/`/g, "'");
-
-  // Remove excessive newlines that could break prompt structure
-  sanitized = sanitized.replace(/\n{4,}/g, "\n\n\n");
-
-  // Add truncation notice if content was cut
-  if (input.length > maxLength) {
-    sanitized += "\n\n[Content truncated for security]";
-  }
-
-  return sanitized;
-}
-
 /**
  * Initialize Bedrock client with AWS credentials
  */

scripts/detect_duplicates.ts

@@ -10,6 +10,7 @@ import {
 } from "@aws-sdk/client-bedrock-runtime";
 import { DuplicateMatch, IssueData } from "./data_models.js";
 import { retryWithBackoff } from "./retry_utils.js";
+import { sanitizePromptInput } from "./sanitize_utils.js";
 
 const MODEL_ID = "us.anthropic.claude-sonnet-4-20250514-v1:0";
 const SIMILARITY_THRESHOLD = 0.8;
@@ -32,50 +33,6 @@ const EXCLUDE_LABELS = new Set<string>([
   "duplicate",
 ]);
 
-/**
- * Sanitize user input to prevent prompt injection attacks
- */
-function sanitizePromptInput(input: string, maxLength: number): string {
-  if (!input) {
-    return "";
-  }
-
-  // Truncate to maximum length
-  let sanitized = input.substring(0, maxLength);
-
-  // Remove potential prompt injection patterns
-  // These patterns could be used to manipulate the AI's behavior
-  const dangerousPatterns = [
-    /ignore\s+(all\s+)?(previous|above|prior)\s+instructions?/gi,
-    /disregard\s+(all\s+)?(previous|above|prior)\s+instructions?/gi,
-    /forget\s+(all\s+)?(previous|above|prior)\s+instructions?/gi,
-    /new\s+instructions?:/gi,
-    /system\s*:/gi,
-    /assistant\s*:/gi,
-    /\[SYSTEM\]/gi,
-    /\[ASSISTANT\]/gi,
-    /\<\|im_start\|\>/gi,
-    /\<\|im_end\|\>/gi,
-  ];
-
-  for (const pattern of dangerousPatterns) {
-    sanitized = sanitized.replace(pattern, "[REDACTED]");
-  }
-
-  // Escape backticks that could break JSON formatting
-  sanitized = sanitized.replace(/`/g, "'");
-
-  // Remove excessive newlines that could break prompt structure
-  sanitized = sanitized.replace(/\n{4,}/g, "\n\n\n");
-
-  // Add truncation notice if content was cut
-  if (input.length > maxLength) {
-    sanitized += "\n\n[Content truncated for security]";
-  }
-
-  return sanitized;
-}
-
 /**
  * Fetch existing open issues from repository as duplicate candidates.
  *

scripts/sanitize_utils.ts

@@ -0,0 +1,58 @@
+/**
+ * Shared input-sanitization utilities.
+ *
+ * Centralizes prompt-input sanitization so every module that builds a model
+ * prompt relies on a single, consistent implementation instead of its own copy.
+ */
+
+/**
+ * Sanitize untrusted user input before embedding it in a model prompt.
+ *
+ * Truncates the input to `maxLength`, redacts a set of known
+ * instruction-override patterns, replaces backticks that could break JSON
+ * formatting, and collapses runs of excessive newlines. A truncation notice is
+ * appended when the original input exceeded `maxLength`.
+ *
+ * @param input - Raw, untrusted text (for example an issue title or body).
+ * @param maxLength - Maximum number of characters to keep from `input`.
+ * @returns The sanitized string, or an empty string when `input` is falsy.
+ */
+export function sanitizePromptInput(input: string, maxLength: number): string {
+  if (!input) {
+    return "";
+  }
+
+  // Truncate to maximum length
+  let sanitized = input.substring(0, maxLength);
+
+  // Remove potential prompt injection patterns
+  const dangerousPatterns = [
+    /ignore\s+(all\s+)?(previous|above|prior)\s+instructions?/gi,
+    /disregard\s+(all\s+)?(previous|above|prior)\s+instructions?/gi,
+    /forget\s+(all\s+)?(previous|above|prior)\s+instructions?/gi,
+    /new\s+instructions?:/gi,
+    /system\s*:/gi,
+    /assistant\s*:/gi,
+    /\[SYSTEM\]/gi,
+    /\[ASSISTANT\]/gi,
+    /\<\|im_start\|\>/gi,
+    /\<\|im_end\|\>/gi,
+  ];
+
+  for (const pattern of dangerousPatterns) {
+    sanitized = sanitized.replace(pattern, "[REDACTED]");
+  }
+
+  // Escape backticks that could break JSON formatting
+  sanitized = sanitized.replace(/`/g, "'");
+
+  // Remove excessive newlines that could break prompt structure
+  sanitized = sanitized.replace(/\n{4,}/g, "\n\n\n");
+
+  // Add truncation notice if content was cut
+  if (input.length > maxLength) {
+    sanitized += "\n\n[Content truncated for security]";
+  }
+
+  return sanitized;
+}