diff --git a/CHANGELOG.md b/CHANGELOG.md
index 14eb2b8b59..aef38e994f 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -111,6 +111,18 @@ file is the human-readable summary of what changed and why.
 
 ### Sandbox
 
+- **Creating a sandbox is near-instant and no longer costs 8 GB** — a new
+  sandbox is meant to start as a copy-on-write clone of the shared disk image,
+  costing no space until the guest writes something. On macOS the clone never
+  happened: the copy fell back to writing the whole 8 GB image out byte for
+  byte, filling in the empty space the image leaves sparse. Creating or
+  resetting a sandbox spent that time writing gigabytes, then the guest booted
+  while the disk was still catching up and had to compete for it — which is why
+  a sandbox's first start could crawl and sometimes gave up with "sandbox
+  server did not become healthy" before finishing. The clone now happens for
+  real: creation is immediate, a sandbox costs roughly nothing on disk until it
+  is used, and first start takes about as long as every start after it.
+  ([#293](https://github.com/quickbeard/codev-code/pull/293))
 - **Work sessions reach the model on WSL's mirrored networking** — on Windows,
   a sandbox's model traffic crosses from the WSL guest to a relay on the
   Windows host, and the guest found the host by assuming WSL's default NAT
diff --git a/packages/codev-sandbox/src/vz/provider.ts b/packages/codev-sandbox/src/vz/provider.ts
index 1d3f73f587..4a2288010b 100644
--- a/packages/codev-sandbox/src/vz/provider.ts
+++ b/packages/codev-sandbox/src/vz/provider.ts
@@ -91,9 +91,8 @@ export function createVzProvider(options: VzProviderOptions): SandboxProvider {
       await writeImageMarker(dirs.base, asset)
     }
     // The pristine unpacked disk stays untouched; the instance runs a
-    // copy-on-write clone (APFS clonefile via FICLONE, plain copy elsewhere),
-    // so reset is just re-cloning.
-    await copyFile(join(dirs.image, VZ_DISK_FILE), dirs.disk, constants.COPYFILE_FICLONE)
+    // copy-on-write clone, so reset is just re-cloning.
+    await cloneDisk(join(dirs.image, VZ_DISK_FILE), dirs.disk, platform)
   }
 
   // A custom image is used from where it lies; otherwise the published one
@@ -309,6 +308,29 @@ export function createVzProvider(options: VzProviderOptions): SandboxProvider {
   }
 }
 
+// A copy-on-write clone of the pristine disk, falling back to a byte copy.
+//
+// `copyFile`'s COPYFILE_FICLONE is a no-op on macOS: libuv opens the
+// destination and calls fcopyfile(), and only clonefile(2) — which takes paths
+// and requires the destination not to exist — can share extents. So every
+// provision wrote the whole image out instead: ~9s and a full 8 GiB per
+// sandbox for a disk that should cost nothing, and it expands the source's
+// sparse holes into real zeroes. Worse, the VM boots straight after, so the
+// guest competes with gigabytes of write-back for the disk it is reading —
+// that contention, not the image itself, is what pushed first boot past the
+// 60s start budget. `cp -c` is the syscall; it fails outright rather than
+// silently copying when the clone cannot happen, which is exactly the signal
+// the fallback needs.
+async function cloneDisk(source: string, destination: string, platform: NodeJS.Platform) {
+  await rm(destination, { force: true })
+  if (platform === "darwin") {
+    const cloned = await runCommand("cp", ["-c", source, destination]).catch(() => null)
+    if (cloned?.code === 0) return
+    await rm(destination, { force: true })
+  }
+  await copyFile(source, destination, constants.COPYFILE_FICLONE)
+}
+
 async function defaultMacosVersion() {
   const result = await runCommand("sw_vers", ["-productVersion"]).catch(() => null)
   if (!result || result.code !== 0) return null