fix(logging): redact Telegram tokens across chunk boundaries (#103861)

src/logging/redact.test.ts

@@ -955,6 +955,24 @@ describe("redactSensitiveText", () => {
     }
   });
 
+  it("masks Telegram bot tokens that cross bounded-replacement chunk boundaries", () => {
+    const chunkSize = 16_384;
+    const credential = `123456:${"A".repeat(28)}WXYZ`;
+    const cases = [
+      { token: `bot${credential}`, redacted: "bot123456…WXYZ" },
+      { token: credential, redacted: "123456…WXYZ" },
+    ];
+
+    for (const { token, redacted } of cases) {
+      const tokenStart = chunkSize - 12;
+      const prefix = `${"x".repeat(tokenStart - 1)} `;
+      const suffix = ` ${"y".repeat(chunkSize * 2)}`;
+      expect(redactSensitiveText(`${prefix}${token}${suffix}`, { mode: "tools" })).toBe(
+        `${prefix}${redacted}${suffix}`,
+      );
+    }
+  });
+
   it("does not corrupt base64 blobs that embed token-prefix shapes", () => {
     // Tiny-PNG base64 contains a gAAAA run from zero-filled IHDR bytes; pure-base64-alphabet
     // prefixes must not fire mid-blob or media payloads get mangled.

src/logging/redact.ts

@@ -127,12 +127,18 @@ const STANDALONE_ASSIGNMENT_REDACT_PATTERN = String.raw`(^|[\s,;])(?:${STANDALON
 // data-URL media is never corrupted while tokens in URL paths or assignments still redact.
 const BASE64_SAFE_TOKEN_BOUNDARY = String.raw`(^|[^A-Za-z0-9])(?<!;base64,[A-Za-z0-9+/=]*)`;
 const IDENTIFIER_SAFE_TOKEN_BOUNDARY = String.raw`(^|[^A-Za-z0-9_])`;
+const TELEGRAM_BOT_TOKEN_REDACT_PATTERN = String.raw`\bbot(\d{6,}:[A-Za-z0-9_-]{20,})\b`;
+const TELEGRAM_TOKEN_REDACT_PATTERN = String.raw`\b(\d{6,}:[A-Za-z0-9_-]{20,})\b`;
 const SHELL_REFERENCE_PRESERVING_PATTERN_SOURCES = new Set([
   ENV_ASSIGNMENT_REDACT_PATTERN,
   ESCAPED_ENV_ASSIGNMENT_REDACT_PATTERN,
   STANDALONE_ASSIGNMENT_QUOTED_REDACT_PATTERN,
   STANDALONE_ASSIGNMENT_REDACT_PATTERN,
 ]);
+const CHUNK_UNSAFE_PATTERN_SOURCES = new Set([
+  TELEGRAM_BOT_TOKEN_REDACT_PATTERN,
+  TELEGRAM_TOKEN_REDACT_PATTERN,
+]);
 const shellReferencePreservingPatterns = new WeakSet<RegExp>();
 // Patterns whose left-context assertions or complete token can cross a chunk boundary must run
 // against the full string; chunking can invent a `^` boundary or split the secret itself.
@@ -251,8 +257,8 @@ const DEFAULT_REDACT_PATTERNS: string[] = [
   String.raw`(api_org_[A-Za-z0-9]{20,})`,
   String.raw`(r8_[A-Za-z0-9]{10,})`,
   // Telegram Bot API URLs embed the token as `/bot<token>/...` (no word-boundary before digits).
-  String.raw`\bbot(\d{6,}:[A-Za-z0-9_-]{20,})\b`,
-  String.raw`\b(\d{6,}:[A-Za-z0-9_-]{20,})\b`,
+  TELEGRAM_BOT_TOKEN_REDACT_PATTERN,
+  TELEGRAM_TOKEN_REDACT_PATTERN,
 ];
 let defaultResolvedPatterns: RegExp[] | undefined;
 
@@ -323,7 +329,9 @@ function parsePattern(raw: RedactPattern): RegExp | null {
   if (
     pattern &&
     typeof raw === "string" &&
-    (raw.startsWith(BASE64_SAFE_TOKEN_BOUNDARY) || raw.startsWith(IDENTIFIER_SAFE_TOKEN_BOUNDARY))
+    (raw.startsWith(BASE64_SAFE_TOKEN_BOUNDARY) ||
+      raw.startsWith(IDENTIFIER_SAFE_TOKEN_BOUNDARY) ||
+      CHUNK_UNSAFE_PATTERN_SOURCES.has(raw))
   ) {
     chunkUnsafePatterns.add(pattern);
   }