Containers, with the boundary visible.
wasm-vm is not Docker Engine in a tab. The current work is an OCI image pipeline plus a
small guest-side runner named wvrun. The page separates what has been
exercised from what is still a boot-gated capstone.
Public-demo caveat. The default public manifest is the small busybox initramfs. It does not imply that the deployed page includes the local-only Alpine container bundles or that the container UI is a live Docker daemon.
What the pipeline does
The native CLI can turn a verified OCI layout into a bundle with
wasm-vm oci unpack <layout> --out <bundle>. The bundle contains
a root filesystem, run.json, and flat configuration files so the guest
runner does not need a JSON parser.
wasm-vm oci unpack <layout> --out <bundle>
wvrun <bundle>
wvrun --interactive <bundle>Kernel capabilities that were audited
The pinned riscv kernel configuration enables the primitives needed by the current runner shape: PID, mount, network, UTS, user, and IPC namespaces; cgroup v2 controls; overlayfs; tmpfs; veth/bridge networking; and loop devices. The in-guest smoke test is important because a config symbol alone is not proof that the emulator implements the corresponding syscall and device path.
| Capability | Used for | Current proof |
|---|---|---|
| Namespaces | PID, mount, UTS, IPC, network, and user isolation. | Kernel audit plus in-guest smoke coverage. |
| cgroup v2 | Memory, PID, and CPU resource boundaries. | Memory limit/OOM event is checked by the smoke test. |
| overlayfs + tmpfs | Read-only bundle lower layer and writable upper layer. | Override and whiteout behavior are exercised. |
| veth + bridge | Guest-side container network plumbing. | Cross-network-namespace ping is exercised in the audit. |
| seccomp | Future runc-style syscall filtering. | Config is enabled; the current wvrun runner does not install the filter yet. |
What wvrun actually provides
The runner creates an overlay, enters PID/mount/UTS/IPC namespaces, mounts fresh
/proc and /dev, pivots into the bundle root, applies the bundle's
argv/environment/cwd/user, and executes the image command with exit-code passthrough.
Its lifecycle surface also includes detached runs, listing, logs, stop/remove, and
exec into a running container.
wvrun run -d --name c1 <bundle>
wvrun ps
wvrun logs c1
wvrun stop c1
wvrun rm c1
wvrun exec -it c1 shThose commands describe the guest runner's checked-in interface. They are not a promise that every public page has a container-capable image or that all Docker CLI behavior is supported.
Image matrix: architecture, not service uptime
The current matrix records 9 of 9 surveyed riscv64 images passing digest-verified unpacking, entrypoint resolution, and whole-rootfs ELF purity: Alpine, BusyBox, nginx, httpd, Caddy, HAProxy, Redis, Memcached, and Postgres. This is a static pipeline result. It does not say that all nine have been booted and served successfully inside the emulator.
Read the result correctly. “9/9 passed” means the image bundle is structurally runnable and every ELF is riscv64 according to the matrix checks. The boot-and-service leg remains the E3.5-T05 capstone.
Current gaps
- The public busybox image does not ship
wvrun; the container-capable Alpine image is a local-only path. - Seccomp is enabled in the kernel configuration but is not yet installed by the current runner.
- The full booted image/service matrix is deferred to the E3.5-T05 capstone and should not be inferred from the native fast path.
- Unsupported architectures are rejected by the image validation path rather than translated into riscv64.
Source records
docs/containers.md— kernel audit, smoke test, and runner contract.docs/container-matrix.md— 9-image static matrix and adversarial self-test.E3.5-T05— the remaining boot-gated acceptance.tools/guest— runner and guest smoke scripts.