diff --git a/.github/OPENCLAW_CI.md b/.github/OPENCLAW_CI.md
index 5307e4898df8..bd15c5777c7e 100644
--- a/.github/OPENCLAW_CI.md
+++ b/.github/OPENCLAW_CI.md
@@ -84,6 +84,12 @@ test/js/node/process/process.test.js
 test/js/bun/sqlite/column-types.test.js
 ```
 
+The `vm.Script` leak regression checks live `Script` cells after collection and
+collects between allocation batches. It measures RSS growth over all 10,000 scripts
+after a separate 5,000-script cache and allocator warmup. It retains the 200 MiB release and 700 MiB ASAN limits; a deliberately
+retained-script control must fail the live-cell assertion when qualifying changes
+to this guard.
+
 ## Windows release qualification
 
 The separate [`openclaw-release.yml`](workflows/openclaw-release.yml) keeps
diff --git a/CHANGELOG.md b/CHANGELOG.md
index a49d8a73e407..f50d50657c91 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -212,3 +212,5 @@
 - Publish Darwin/Linux prereleases while Windows signing is unconfigured; require an explicit signed-Windows release switch before including either Windows architecture, while preserving test-only Windows CI.
 
 - Avoid requesting executable access when renaming files on Windows, preventing unnecessary synchronous antivirus scans during compile-cache publication.
+
+- Check live `vm.Script` counts and full-workload RSS growth after warmup in the leak regression, preserving the 200 MiB release and 700 MiB ASAN RSS limits. Adapts [oven-sh/bun#42474](https://github.com/oven-sh/bun/pull/42474); thanks @robobun!
diff --git a/test/js/node/vm/script-leak.test.ts b/test/js/node/vm/script-leak.test.ts
index 586ec3d88b99..c77a091eda9b 100644
--- a/test/js/node/vm/script-leak.test.ts
+++ b/test/js/node/vm/script-leak.test.ts
@@ -1,30 +1,41 @@
-const vm = require("vm");
-const { describe, it, expect } = require("bun:test");
-const { isASAN, rss } = require("harness");
+import { heapStats } from "bun:jsc";
+import { describe, expect, it } from "bun:test";
+import { isASAN, rss } from "harness";
+import vm from "node:vm";
+
+const iterations = 10_000;
+const warmupIterations = 5_000;
+const batch = 500;
 
 describe("vm.Script", () => {
   it("shouldn't leak memory", () => {
-    const initialUsage = rss();
-
-    {
-      const source = `/*\n${Buffer.alloc(10000, " * aaaaa\n").toString("utf8")}\n*/ Buffer.alloc(10, 'hello');`;
-
-      function go(i) {
-        const script = new vm.Script(source + "//" + i);
-        script.runInThisContext();
-      }
+    Bun.gc(true);
+    const initialCount = heapStats().objectTypeCounts.Script ?? 0;
+    let initialUsage = 0;
+    // The measured source payload exceeds the RSS bound if native sources are retained.
+    const source = `/*\n${Buffer.alloc(32_768, " * aaaaa\n").toString("utf8")}\n*/ Buffer.alloc(10, 'hello').toString();`;
 
-      for (let i = 0; i < 10000; ++i) {
-        go(i);
-      }
+    let result;
+    function go(i) {
+      const script = new vm.Script(source + "//" + i);
+      result = script.runInThisContext();
     }
 
+    for (let i = 0; i < warmupIterations + iterations; ++i) {
+      go(i);
+      // Bound temporary garbage before it inflates the allocator's resident pages.
+      if ((i + 1) % batch === 0) Bun.gc(true);
+      // Measure all 10,000 scripts after a separate cache and allocator warmup.
+      if (i + 1 === warmupIterations) initialUsage = rss();
+    }
+    expect(result).toBe("hellohello");
     Bun.gc(true);
 
     const finalUsage = rss();
+    const finalCount = heapStats().objectTypeCounts.Script ?? 0;
     const megabytes = Math.round(((finalUsage - initialUsage) / 1024 / 1024) * 100) / 100;
-    // ASAN's quarantine retains freed allocations (default 256 MB) so RSS
-    // deltas run far higher under bun-asan; widen the threshold there.
+    expect(finalCount).toBeLessThanOrEqual(initialCount + 10);
+    // ASAN's quarantine retains freed allocations (default 256 MB).
     expect(megabytes).toBeLessThan(isASAN ? 700 : 200);
   });
 });