diff --git a/src/compile_cache_zstd.dict b/src/compile_cache_zstd.dict
index b64455d45b1d82..ef9c5cb52a23bf 100644
Binary files a/src/compile_cache_zstd.dict and b/src/compile_cache_zstd.dict differ
diff --git a/tools/train_compile_cache_dict.mjs b/tools/train_compile_cache_dict.mjs
new file mode 100644
index 00000000000000..bce24bdeede3bf
--- /dev/null
+++ b/tools/train_compile_cache_dict.mjs
@@ -0,0 +1,168 @@
+#!/usr/bin/env node
+// =============================================================================
+// train_compile_cache_dict.mjs
+//
+// Maintainer tool that regenerates src/compile_cache_zstd.dict, the zstd
+// dictionary embedded in the node binary and used to shrink compile-cache
+// entries on disk (see src/compile_cache.cc).
+//
+// -----------------------------------------------------------------------------
+// What it does
+// -----------------------------------------------------------------------------
+// The node compile cache stores V8 code caches (the bytecode/metadata blob V8
+// produces when it compiles a module) so later runs can skip recompilation.
+// These blobs share a lot of structure across files, so we zstd-compress them
+// with a trained dictionary to cut the cache's on-disk footprint.
+//
+// This script builds that dictionary end to end:
+//
+//   1. Walks a fixed in-tree corpus (CORPUS below) in sorted order.
+//   2. For each .js file, harvests a V8 code cache via vm.compileFunction with
+//      produceCachedData — the same shape the CommonJS loader produces at
+//      runtime — and writes each blob to a temp directory.
+//   3. Feeds all the harvested blobs to `zstd --train`, capping the result at
+//      MAXDICT (16 KiB) bytes.
+//   4. Overwrites src/compile_cache_zstd.dict with the trained dictionary.
+//
+// The shipped src/compile_cache_zstd.dict is the source of truth; this script
+// exists to document and reproduce exactly how that file was made, so a future
+// maintainer can regenerate it (e.g. after a V8/corpus change) and review the
+// resulting diff.
+//
+// -----------------------------------------------------------------------------
+// Usage
+// -----------------------------------------------------------------------------
+//   node tools/train_compile_cache_dict.mjs
+//
+// Run from anywhere (paths are resolved relative to this file). The output is
+// written in place to src/compile_cache_zstd.dict; inspect the diff afterwards
+// and commit it if it looks right. Progress is printed to stderr.
+//
+// Prerequisites:
+//   * the `zstd` CLI on PATH, version REQUIRED_ZSTD (matching deps/zstd).
+//   * a built node on PATH (this is the node you invoke it with).
+//
+// Note: --predictable is added automatically (see below) — you do not need to
+// pass it yourself.
+//
+// -----------------------------------------------------------------------------
+// Reproducibility
+// -----------------------------------------------------------------------------
+// The output is byte-for-byte stable only if all of the following are pinned:
+//
+//   * node is run with --predictable. V8 otherwise randomizes the string hash
+//     seed per process, and that seed leaks into vm.compileFunction cachedData,
+//     so every harvest (and therefore every trained dictionary) would differ
+//     run-to-run even on the same machine. This script re-executes itself with
+//     --predictable if needed.
+//   * the corpus and its order are fixed (CORPUS below, walked sorted).
+//   * the node build is fixed: cachedData embeds the V8 version and build
+//     flags, so a different node produces a different (still valid) dictionary.
+//   * the zstd CLI matches deps/zstd (currently 1.5.7). ZDICT training defaults
+//     change between zstd releases; this script refuses to run on a mismatch.
+//
+// Training on --predictable caches is fine even though the runtime consumes
+// non-predictable caches: the dictionary only supplies shared substrings, and
+// Persist() keeps min(plain, dict) per entry, so a less-than-ideal dictionary
+// can never make any entry larger.
+// =============================================================================
+
+import { spawnSync } from 'node:child_process';
+import { readFileSync, writeFileSync, mkdtempSync, readdirSync,
+         rmSync } from 'node:fs';
+import { join, relative, dirname } from 'node:path';
+import { tmpdir } from 'node:os';
+import { fileURLToPath } from 'node:url';
+
+const ROOT = join(dirname(fileURLToPath(import.meta.url)), '..');
+const OUT = join(ROOT, 'src', 'compile_cache_zstd.dict');
+const MAXDICT = 16384;
+const REQUIRED_ZSTD = '1.5.7';      // keep in sync with deps/zstd/lib/zstd.h
+
+// Fixed corpus, relative to the repo root. Chosen to be diverse, always
+// present in a checkout, and disjoint from the held-out measurement corpora
+// (e.g. test/parallel) used to report size numbers.
+const CORPUS = ['lib', 'tools', 'deps/npm/node_modules'];
+
+// V8 randomizes the hash seed per process; re-exec under --predictable so the
+// harvested caches — and the trained dictionary — are deterministic.
+if (!process.execArgv.includes('--predictable')) {
+  const r = spawnSync(process.execPath,
+    ['--predictable', fileURLToPath(import.meta.url), ...process.argv.slice(2)],
+    { stdio: 'inherit' });
+  process.exit(r.status ?? 1);
+}
+
+function checkZstd() {
+  const r = spawnSync('zstd', ['--version'], { encoding: 'utf8' });
+  if (r.status !== 0) {
+    console.error('error: zstd CLI not found on PATH'); process.exit(1);
+  }
+  const m = r.stdout.match(/v(\d+\.\d+\.\d+)/);
+  if (!m || m[1] !== REQUIRED_ZSTD) {
+    console.error(`error: zstd ${REQUIRED_ZSTD} required (matching deps/zstd), ` +
+                  `found ${m ? m[1] : 'unknown'}`);
+    process.exit(1);
+  }
+}
+
+const PARAMS = ['exports', 'require', 'module', '__filename', '__dirname'];
+
+function* walk(dir) {
+  let ents;
+  try { ents = readdirSync(dir, { withFileTypes: true }); } catch { return; }
+  for (const e of ents.sort((a, b) => (a.name < b.name ? -1 : 1))) {
+    const p = join(dir, e.name);
+    if (e.isDirectory()) yield* walk(p);
+    else if (e.isFile() && p.endsWith('.js')) yield p;
+  }
+}
+
+async function main() {
+  checkZstd();
+  const { default: vm } = await import('node:vm');
+
+  const files = [];
+  for (const root of CORPUS) for (const f of walk(join(ROOT, root))) files.push(f);
+  files.sort();
+
+  const samples = mkdtempSync(join(tmpdir(), 'cc-dict-'));
+  const sampleFiles = [];
+  let ok = 0;
+  let r;
+  try {
+    for (const f of files) {
+      let code;
+      try { code = readFileSync(f, 'utf8'); } catch { continue; }
+      try {
+        const fn = vm.compileFunction(code, PARAMS,
+          { filename: f, produceCachedData: true });
+        const cd = fn.cachedData;
+        if (cd && cd.length > 0) {
+          const name = relative(ROOT, f).replace(/[\\/]/g, '_') + '.cache';
+          const out = join(samples, name);
+          writeFileSync(out, cd);
+          sampleFiles.push(out);
+          ok++;
+        }
+      } catch { /* skip modules V8 can't compile standalone */ }
+    }
+    sampleFiles.sort();
+    console.error(`harvested ${ok}/${files.length} code caches from ` +
+                  `${CORPUS.join(', ')}`);
+
+    // One sample path per file is spread on argv; the fixed corpus (~1.4k
+    // files, a few hundred KB of paths) stays well under ARG_MAX.
+    r = spawnSync('zstd',
+      ['--train', ...sampleFiles, `--maxdict=${MAXDICT}`, '-f', '-o', OUT],
+      { stdio: ['ignore', 'ignore', 'inherit'] });
+  } finally {
+    rmSync(samples, { recursive: true, force: true });
+  }
+  if (r.status !== 0) { console.error('error: zstd --train failed'); process.exit(1); }
+
+  const size = readFileSync(OUT).length;
+  console.error(`wrote ${relative(ROOT, OUT)} (${size} bytes)`);
+}
+
+main().catch((e) => { console.error(e); process.exit(1); });