quico 0.4.0 → 0.8.0

package/src/ja4h.js

@@ -0,0 +1,256 @@
+/*
+ * quico: a full HTTP stack in JavaScript
+ * Copyright 2025 colocohen
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     https://github.com/colocohen/quico
+ */
+
+
+
+import { sha12 } from './http_common.js';
+
+
+// ============================================================
+//  JA4H — HTTP client fingerprint.
+//
+//  The companion to the TLS fingerprints lemon-tls produces. JA4 says which
+//  TLS stack built the ClientHello; JA4H says which HTTP client sent the
+//  request. Together they catch a class of thing neither finds alone: a
+//  request whose TLS says "Chrome 131" and whose headers say "python-requests"
+//  is not a browser, however convincing its User-Agent.
+//
+//  Format:  a_b_c_d
+//
+//    a  ge20cn11enus   method, version, cookie, referer, header count, lang
+//    b  3a41b52c00e1   hash of the header NAMES, in the order sent
+//    c  000000000000   hash of the cookie names, sorted
+//    d  000000000000   hash of the cookie name=value pairs, sorted
+//
+//  The sections are separately useful, which is the point of the a_b_c_d
+//  shape. FoxIO's own description: JA4H_c identifies the SITE (every Plex
+//  server sets the same cookie names, so every client of one produces the same
+//  _c), while JA4H_d identifies the USER (their values differ). That makes _d
+//  a per-user handle that contains no personal data — a tracking key that is
+//  a hash of a hash, rather than a logged identity.
+//
+//  A caution worth stating plainly: _d IS a user identifier, GDPR-friendly in
+//  form or not. Log it if you need per-user correlation, and think before you
+//  keep it longer than the session it describes.
+//
+//  The published spec (JA4H.md) is 278 bytes and documents ONLY the
+//  header-count rule. Everything else here is reconstructed from FoxIO's
+//  reference implementations — chiefly the F5 iRule, which carries their
+//  copyright — and then verified against a published fingerprint rather than
+//  trusted.
+//
+//  That verification mattered, because FoxIO's own implementations disagree
+//  about _b: the Rust CLI documents `--original-order` as DISABLING the
+//  sorting of headers, implying it sorts by default, while the iRule appends
+//  header names in the order received and never sorts them.
+//
+//  The disagreement was settled by arithmetic. The Sliver C2 analysis
+//  publishes po11cn050000_bb52516416a2_eb49a3237520 for a request whose five
+//  counted headers are Host, User-Agent, Content-Length,
+//  Upgrade-Insecure-Requests, Accept-Encoding. Hashing every plausible
+//  variant of that list gives exactly one match:
+//
+//      original order, original case  ->  bb52516416a2   <-- published value
+//      original order, lowercased     ->  4637073cda60
+//      sorted, original case          ->  a56dc1e0e2e6
+//      sorted, lowercased             ->  7c182098df7a
+//
+//  So: order as received, case as received. Which is also the reading that
+//  makes sense — ORDER is most of the signal, since browsers emit a stable
+//  header order and HTTP libraries emit a different, equally stable one, and
+//  sorting would discard exactly that.
+// ============================================================
+
+var EMPTY_HASH = '000000000000';
+
+
+// Two lowercase characters. Anything not in this list is "xx" — which is
+// itself a signal, since browsers only ever send these.
+var METHOD_CODES = {
+  GET: 'ge', POST: 'po', HEAD: 'he', PUT: 'pu', DELETE: 'de',
+  CONNECT: 'co', OPTIONS: 'op', TRACE: 'tr', PATCH: 'pa'
+};
+
+function methodCode(method) {
+  var m = String(method || '').toUpperCase();
+  return METHOD_CODES[m] || 'xx';
+}
+
+function versionCode(httpVersion) {
+  var v = String(httpVersion || '');
+  if (v === '1.0' || v === '10') return '10';
+  if (v === '1.1' || v === '11') return '11';
+  if (v === '2.0' || v === '2' || v === '20') return '20';
+  if (v === '3.0' || v === '3' || v === '30') return '30';
+  return '00';
+}
+
+/**
+ * Four characters from Accept-Language, or "0000" when absent.
+ *
+ * "en-US,en;q=0.9,he;q=0.8" -> "enus": take the first language, drop the
+ * q-value, remove the hyphen, lowercase, then truncate or zero-pad to four.
+ *
+ * Four characters is deliberately coarse. It separates populations — a client
+ * with no Accept-Language at all stands out sharply, and most HTTP libraries
+ * send none — without turning a language preference into an identifier.
+ */
+function languageCode(acceptLanguage) {
+  if (!acceptLanguage) return '0000';
+
+  // Bounded before anything else. Accept-Language is attacker-controlled and
+  // unbounded; the reference implementation truncates to 64 bytes first for
+  // exactly that reason.
+  var v = String(acceptLanguage).slice(0, 64);
+
+  // Hyphens removed and semicolons turned into commas BEFORE splitting, so
+  // that the q-value separator and the language separator are treated alike.
+  // "en-US,en;q=0.9" -> "enus,en,q=09" -> first token "enus".
+  v = v.replace(/-/g, '').replace(/;/g, ',').toLowerCase();
+  var first = v.split(',')[0].slice(0, 4);
+  if (first.length === 0) return '0000';
+
+  // Padded by OVERWRITING the leading characters of "0000", so a 2-character
+  // language becomes "he00" rather than "00he".
+  return first + '0'.repeat(4 - first.length);
+}
+
+/** Cookie names and values, in the order sent. */
+function parseCookies(cookieHeader) {
+  var out = [];
+  if (!cookieHeader) return out;
+  var parts = String(cookieHeader).split(';');
+  for (var i = 0; i < parts.length; i++) {
+    var p = parts[i].trim();
+    if (!p) continue;
+    var eq = p.indexOf('=');
+    if (eq < 0) out.push({ name: p, value: '' });
+    else out.push({ name: p.slice(0, eq).trim(), value: p.slice(eq + 1).trim() });
+  }
+  return out;
+}
+
+
+/**
+ * Compute JA4H for one request.
+ *
+ * Pass a quico (or node:http) IncomingMessage, or a plain object with
+ * `method`, `httpVersion`, `headers` and optionally `rawHeaders`.
+ *
+ * rawHeaders matters and is used when present: it preserves the ORDER the
+ * client sent, and order is most of what distinguishes a browser from a
+ * library. A `headers` object cannot preserve it reliably — repeated fields
+ * are folded and insertion order is an implementation detail — so passing only
+ * `headers` still produces a valid fingerprint, just a weaker one.
+ *
+ * Returns { hash, raw, a, b, c, d } so a caller can match on the whole thing
+ * or on any section: `_ab` identifies the client software, `_c` the site, `_d`
+ * the user.
+ */
+function ja4h(req) {
+  var headers = (req && req.headers) || {};
+  var method = req && (req.method || headers[':method']);
+  var version = req && req.httpVersion;
+
+  // Header names in the order the client sent them.
+  // ORIGINAL CASE, not lowercased.
+  //
+  // FoxIO's own iRule appends `${hname}` — the name as sent — and lowercases
+  // only for its comparisons. Casing is real signal: HTTP/1.1 clients choose
+  // it freely and choose differently ("User-Agent" from a browser,
+  // "user-agent" from many libraries), so folding it throws away a
+  // distinguishing bit. HTTP/2 and HTTP/3 require lowercase on the wire, so
+  // there is nothing to lose there either way.
+  var names = [];
+  if (req && Array.isArray(req.rawHeaders) && req.rawHeaders.length >= 2) {
+    for (var i = 0; i < req.rawHeaders.length; i += 2) {
+      names.push(String(req.rawHeaders[i]));
+    }
+  } else {
+    for (var k in headers) {
+      if (Object.prototype.hasOwnProperty.call(headers, k)) names.push(String(k));
+    }
+  }
+
+  // Pseudo-headers are HTTP/2 and HTTP/3 framing, not client-chosen fields,
+  // and counting them would make the same client fingerprint differently
+  // depending on which version carried the request — the opposite of what a
+  // client fingerprint is for.
+  names = names.filter(function (n) { return n.charAt(0) !== ':'; });
+
+  function lower(n) { return String(n).toLowerCase(); }
+  var lowered = names.map(lower);
+
+  var hasCookie = lowered.indexOf('cookie') !== -1;
+  // "referer" only — the misspelling is the one in the HTTP spec, and
+  // "referrer" is a different header that a client would have invented.
+  var hasReferer = lowered.indexOf('referer') !== -1;
+
+  // Cookie and Referer are excluded from both the count and the hash: they are
+  // properties of the SITE and the navigation, not of the client. Their
+  // presence is already recorded as the c/r characters, and the cookies
+  // themselves get their own sections.
+  // "cookie*", not just "cookie": the reference implementation matches the
+  // prefix, which also excludes Cookie2 and any other cookie-carrying variant.
+  var counted = names.filter(function (n) {
+    var l = lower(n);
+    return l.indexOf('cookie') !== 0 && l !== 'referer';
+  });
+
+  var count = counted.length;
+  if (count > 99) count = 99;
+  var countStr = (count < 10 ? '0' : '') + count;
+
+  var a = methodCode(method) +
+          versionCode(version) +
+          (hasCookie ? 'c' : 'n') +
+          (hasReferer ? 'r' : 'n') +
+          countStr +
+          languageCode(headers['accept-language']);
+
+  // Order preserved, not sorted. Browsers emit a stable header order and HTTP
+  // libraries emit a different, equally stable one, so the order carries more
+  // signal than the set does.
+  var bRaw = counted.join(',');
+  var b = counted.length ? sha12(bRaw) : EMPTY_HASH;
+
+  // Keyed off hasCookie — the same flag that produced the 'c' character —
+  // rather than off whether any cookie parsed. A Cookie header that is present
+  // but empty still means "this client sends cookies", and the two sections
+  // must agree with the character or the fingerprint contradicts itself.
+  var cookies = hasCookie ? parseCookies(headers['cookie']) : [];
+
+  // Sorted, unlike the header names: cookie order is not under the client's
+  // control in the way header order is, so sorting removes noise rather than
+  // signal.
+  var cookieNames = cookies.map(function (x) { return x.name; }).sort();
+  var cRaw = cookieNames.join(',');
+  var c = hasCookie ? sha12(cRaw) : EMPTY_HASH;
+
+  var cookiePairs = cookies.map(function (x) { return x.name + '=' + x.value; }).sort();
+  var dRaw = cookiePairs.join(',');
+  var d = hasCookie ? sha12(dRaw) : EMPTY_HASH;
+
+  return {
+    hash: a + '_' + b + '_' + c + '_' + d,
+
+    // The raw form deliberately STOPS at the cookie names and omits their
+    // values, which is what the reference implementation does and for a reason
+    // worth repeating: cookie values are session tokens. The _d hash above
+    // already captures that dimension, so printing the values into a log adds
+    // nothing but a credential leak.
+    raw: a + '_' + bRaw + '_' + cRaw,
+    a: a, b: b, c: c, d: d
+  };
+}
+
+
+export { ja4h, languageCode, methodCode, versionCode };