diff --git a/packages/server/test/helpers.ts b/packages/server/test/helpers.ts
index e77c1b2..b8b9b2a 100644
--- a/packages/server/test/helpers.ts
+++ b/packages/server/test/helpers.ts
@@ -66,7 +66,11 @@ export async function createTestApp(options: TestAppOptions = {}): Promise<TestA
     cleanup: async () => {
       deps.channels.dispose();
       deps.db.close();
-      await fs.rm(root, { recursive: true, force: true });
+      // maxRetries: Windows can report ENOTEMPTY/EBUSY while handles from the test's own
+      // just-closed files (SQLite, trace writers) are still being released — Node's rm
+      // retries these codes with a delay. A plain rm was the top cause of ci-windows
+      // cascades (cleanup throws → hook timeout → "database is not open" in later files).
+      await fs.rm(root, { recursive: true, force: true, maxRetries: 10, retryDelay: 100 });
     },
   };
 }
diff --git a/packages/server/vitest.config.ts b/packages/server/vitest.config.ts
new file mode 100644
index 0000000..f16ba7e
--- /dev/null
+++ b/packages/server/vitest.config.ts
@@ -0,0 +1,19 @@
+/**
+ * Vitest config for the server. Mirrors core's platform-aware deadline rationale: these
+ * tests build full app instances over real SQLite files and Agent State trees, and the
+ * Windows CI runners' I/O variance (Defender first-touch scans, slow handle release)
+ * pushes individually fast tests and cleanup hooks past the 5s/10s defaults — a
+ * different file on each run. Larger deadlines change nothing for passing tests; POSIX
+ * keeps the defaults to fail fast during local development. The hook timeout matters
+ * here specifically: `cleanup()` removes the whole test root, and on win32 that rm can
+ * legitimately take retries (see helpers.ts).
+ */
+import { defineConfig } from "vitest/config";
+
+export default defineConfig({
+  test: {
+    environment: "node",
+    testTimeout: process.platform === "win32" ? 30_000 : 5_000,
+    hookTimeout: process.platform === "win32" ? 30_000 : 10_000,
+  },
+});