@mnemonik/shared 7.97.2 → 7.98.0

package/src/secretPatterns.ts

@@ -2,26 +2,26 @@
  * Single source of truth for secret-redaction patterns.
  *
  * Used by:
- * - packages/shared CodeScanner — scrubs chunk content before computing
+ * - packages/shared CodeScanner - scrubs chunk content before computing
  *   contentHash, so daemon ships scrubbed content (correct hash for
  *   server-side cache dedup).
- * - server /api/v1/scan/push handler — re-applies scrub as defense in
- *   depth (idempotent — already-scrubbed content stays the same), so
+ * - server /api/v1/scan/push handler - re-applies scrub as defense in
+ *   depth (idempotent - already-scrubbed content stays the same), so
  *   older daemons or compromised daemons can't leak secrets through us.
- * - server GitMiner — scrubs commit messages before storing as memories.
+ * - server GitMiner - scrubs commit messages before storing as memories.
  *
  * Patterns target high-confidence credential shapes:
  *   1. key=value style: api_key, secret, token, password, credential, auth
  *   2. Stripe-style sk_live_/pk_test_ keys
  *   3. GitHub personal access tokens (ghp_ prefix, exact 36 chars)
  *   4. GitLab personal access tokens (glpat- prefix, 20+ chars)
- *   5. PEM private keys — whole block, header through footer
+ *   5. PEM private keys - whole block, header through footer
  *   6. Provider prefixes: AWS (AKIA/ASIA + labeled secret key), JWT,
  *      Slack (xox*), Google (AIza), Anthropic (sk-ant-), npm (npm_/npms_)
  *   7. Credentials inside scheme://user:pass@host connection strings
  *   8. Authorization: Bearer headers
  *   9. Standalone high-entropy tokens matching no known prefix
- *      (`redactHighEntropyTokens` — heavily guarded, see below)
+ *      (`redactHighEntropyTokens` - heavily guarded, see below)
  *
  * False-positive cost: a few legitimate strings get replaced with the
  * placeholder. False-negative cost: a credential ships to the server and
@@ -32,10 +32,10 @@
  * FIDELITY IS A PEER CONCERN, NOT A ROUNDING ERROR. Over-scrubbing is the
  * same class of failure as under-scrubbing: a pattern that eats prose,
  * identifiers, hashes or paths silently blinds doc-truth's authority
- * extractors — exactly what happened when a `process.env.X` read was
+ * extractors - exactly what happened when a `process.env.X` read was
  * redacted as if it were a literal (see ENV_READ_VALUE_RE below). Every
  * pattern here carries positive AND negative tests in
- * tests/SecretPatterns.test.ts, and the entropy detector is additionally
+ * tests/SecretPatterns.test.ts, and the entropy detector is also
  * guarded against hashes, UUIDs, paths, slugs and identifiers.
  */
 
@@ -45,7 +45,7 @@ export const SECRET_REDACTION_PLACEHOLDER = '[REDACTED]';
 export const SECRET_PATTERNS: ReadonlyArray<RegExp> = [
   /(?:api[_-]?key|secret|token|password|credential|auth)\s*[:=]\s*\S+/gi,
   // Stripe-shape: (sk|pk)_(live|test)_<24+ alphanumerics>. Catches modern
-  // Stripe keys whose body is split by an environment underscore that
+  // Stripe keys whose body is split by an environment separator that
   // breaks the contiguous-alphanum pattern below. Required `live|test`
   // literal prevents false-positives on snake_case identifiers like
   // pkg_install_helper_function_xyz_abc_def.
@@ -53,7 +53,7 @@ export const SECRET_PATTERNS: ReadonlyArray<RegExp> = [
   /(?:sk|pk)[-_][a-zA-Z0-9]{20,}/g,
   /ghp_[a-zA-Z0-9]{36}/g,
   /glpat-[a-zA-Z0-9-]{20,}/g,
-  // PEM private keys. The BLOCK pattern runs first so header→footer (the
+  // PEM private keys. The BLOCK pattern runs first so header->footer (the
   // actual key material) is redacted as one unit; the header-only pattern
   // stays as the fallback for the case a chunk boundary lands between
   // header and footer, where the block never closes inside the chunk.
@@ -82,7 +82,7 @@ export const SECRET_PATTERNS: ReadonlyArray<RegExp> = [
   /\b[a-z][a-z0-9+.-]*:\/\/[^/\s?#]*@/gi,
   // Authorization: Bearer <token>. The token class deliberately excludes
   // `<` and `$`, so documentation placeholders (`Bearer <your-api-key>`,
-  // `Bearer $MNEMONIK_PROXY_TOKEN`) are left intact — redacting those would
+  // `Bearer $MNEMONIK_PROXY_TOKEN`) are left intact - redacting those would
   // destroy the instruction without hiding a secret.
   /Authorization\s*:\s*Bearer\s+[A-Za-z0-9._~+/-]+=*/gi,
 ];
@@ -97,7 +97,7 @@ export const SECRET_PATTERNS: ReadonlyArray<RegExp> = [
 /**
  * A key=value match whose value side is a code REFERENCE to an env read
  * (`process.env.X`, `process.env['X']`, `import.meta.env.X`) contains no
- * literal secret — redacting it destroys information (it blinded the
+ * literal secret - redacting it destroys information (it blinded the
  * doc-truth env_vars authority to every secret-named variable) without
  * protecting anything.
  */
@@ -106,7 +106,7 @@ const ENV_READ_VALUE_RE = /[:=]\s*(?:await\s+)?(?:process\.env[.[]|import\.meta\
 /* ------------------------------------------------------------------ *
  * High-entropy token detector
  *
- * Catches credentials that carry no recognizable provider prefix — the
+ * Catches credentials that carry no recognizable provider prefix - the
  * long random blob in `"webhook": "<40 random chars>"`. A naive entropy
  * threshold cannot do this safely: measured over this repository, a
  * camelCase identifier scores 4.49 bits/char against 4.66 for a real AWS
@@ -114,10 +114,10 @@ const ENV_READ_VALUE_RE = /[:=]\s*(?:await\s+)?(?:process\.env[.[]|import\.meta\
  * below carry the discrimination; entropy is only the final filter.
  *
  * Every candidate must clear ALL of:
- *   1. CONTEXT     — sits immediately after `=`/`:` or an opening quote.
+ *   1. CONTEXT     - sits immediately after `=`/`:` or an opening quote.
  *                    Bare prose positions are never touched, which is what
  *                    keeps the detector out of documentation.
- *   2. LENGTH      — >= 36 characters. The floor started at 32 and was
+ *   2. LENGTH      - >= 36 characters. The floor started at 32 and was
  *                    raised by the claim-yield guard: 32-char opaque
  *                    RESOURCE IDS (Vercel `dpl_`/`prj_`, Stripe object ids)
  *                    are public identifiers, not credentials, and eating
@@ -127,19 +127,19 @@ const ENV_READ_VALUE_RE = /[:=]\s*(?:await\s+)?(?:process\.env[.[]|import\.meta\
  *                    base64 32-byte key 44); shorter non-hex randomness is
  *                    almost always an object id, and anything hex is
  *                    already exempt by rule 4.
- *   3. CHARSET MIX — contains lowercase AND uppercase AND a digit.
- *   4. NOT HEX-ISH — hex after stripping `-`/`_` is a digest or UUID
+ *   3. CHARSET MIX - contains lowercase AND uppercase AND a digit.
+ *   4. NOT HEX-ISH - hex after stripping `-`/`_` is a digest or UUID
  *                    (commit SHAs, contentHash, snippet_hash, project ids).
- *   5. NOT STRUCTURED — if every `/`, `-`, `_` separated segment is <= 16
+ *   5. NOT STRUCTURED - if every `/`, `-`, `_` separated segment is <= 16
  *                    chars the token is a path, slug or SCREAMING_SNAKE
  *                    identifier ("tests/fixtures/docTruth/dogfood-2026-06-06").
- *   6. NOT WORDY   — a lowercase run longer than 5 means natural words
+ *   6. NOT WORDY   - a lowercase run longer than 5 means natural words
  *                    ("...FromCache..."), not random output.
- *   7. ENTROPY     — Shannon entropy >= 4.0 bits/char.
+ *   7. ENTROPY     - Shannon entropy >= 4.0 bits/char.
  *
  * Calibrated by sweeping every file this scanner would chunk across this
  * repository (1426 files): 8 tokens fire, every one of them a credential
- * shape — JWT headers, an AWS example secret key, a leaked Stripe
+ * shape - JWT headers, an AWS example secret key, a leaked Stripe
  * `whsec_`, and this suite's own fixtures. Zero prose, zero paths, zero
  * identifiers, zero resource ids.
  *
@@ -154,14 +154,16 @@ const ENTROPY_MAX_LOWERCASE_RUN = 5;
 
 /**
  * Maximal runs of base64url/base64 characters. `=` is admitted only as
- * trailing padding — allowing it inside the run let a candidate span an
+ * trailing padding - allowing it inside the run let a candidate span an
  * assignment (`HARNESS_PROJECT_ID=<uuid>` matched as one token, defeating
  * the UUID exemption).
  */
-const ENTROPY_CANDIDATE_RE = /[A-Za-z0-9+/_-]{36,}={0,2}/g;
+// Keep the same 36-character floor without an unbounded counted repetition:
+// V8 retains backtracking state for {36,} and overflows on a 10 MiB line.
+const ENTROPY_CANDIDATE_RE = /[A-Za-z0-9+/_-]{35}[A-Za-z0-9+/_-]+={0,2}/g;
 /** Opening quote, or `=`/`:` with optional whitespace and optional quote. */
 const ENTROPY_CONTEXT_RE = /(?:["'`]|[:=]\s*["'`]?)$/;
-/** `==`, `=>`, `!=`, `<=`, `>=`, `+=` … are comparisons, not assignments. */
+/** `==`, `=>`, `!=`, `<=`, `>=`, `+=` ... are comparisons, not assignments. */
 const ENTROPY_OPERATOR_TAIL_RE = /[=!<>+\-*/%&|^]$/;
 const ENTROPY_HEXISH_RE = /^[0-9a-fA-F]+$/;
 const ENTROPY_STRUCTURED_RE = /^[A-Za-z0-9]+(?:[/_-][A-Za-z0-9]+)+$/;
@@ -209,7 +211,7 @@ function isHighEntropySecret(token: string): boolean {
 
 /**
  * Redact standalone high-entropy tokens that no provider pattern claimed.
- * Only the token is replaced — the surrounding key, quote and punctuation
+ * Only the token is replaced - the surrounding key, quote and punctuation
  * survive, so `{"webhook": "<secret>"}` stays valid, readable JSON.
  *
  * Exported for direct testing of the guard behavior; production callers