Skip to content

moca v0.8.0 — microVM sandbox egress

Choose a tag to compare

@pdettori pdettori released this 26 Oct 14:33
d8bc58b

moca v0.8.0 — microVM sandbox egress

The headline is P4.1: the per-Exec Firecracker microVMs behind mocactl sbx can now reach the
outside world — through a shared host proxy, on MOCA's terms. Until now a guest had only lo: any
git clone, package install or API call from a sandbox failed. With this release the guest's
127.0.0.1:3128 is byte-spliced over vsock CID 2:1025 to moca-egress, a per-host forward proxy
that decides, per connection and per request, what may leave (#277).

How egress works

  • The transport (#582): the vsock listener is bound and handed to moca-egress before the jailer
    starts, so a guest can never boot into a listenerless host; a failed Register fails the VM restore.
    Teardown is SIGKILL → wait → Unregister → unlink, on both cleanup paths. The rootfs exports
    HTTPS_PROXY=http://127.0.0.1:3128 and a placeholder git credential, so the usual tooling just
    works. Cloud Hypervisor VMs refuse egress by name.
  • The proxy (#583): identity comes from the accepting socket only — headers never set it, and
    X-Moca-*/Proxy-* are stripped. Per-VM connection and per-session request caps (429), a
    required host allowlist (403 destination_denied, never a timeout), resolve-once with a refusal if
    any answer is private (loopback, RFC 1918, CGNAT, ULA, the host itself — which rules out DNS
    rebinding), opaque CONNECT splices, scrubbed plain-HTTP forwarding, and one JSON audit line per
    decision. The credential seam fails closed: 503 credential_unavailable, never a tunnel.
  • A public conformance suite (moca-egress-conformance, 13 mutation-checked cases) pins the
    policy, and the KVM gate TestGateEgressThroughVsock plus the NoNetworkDevice gate assert the
    properties live on the rig. The guest still has only lo — a direct dial must fail with
    "Network is unreachable".
  • Deploy: deploy/microvm gains moca-egress.service (DynamicUser, empty capability set,
    ProtectSystem=strict) ordered before microvm-worker; setup-microvm.sh installs and seeds it.
    The operator guide is the new "Guest egress" section of
    P4-ON-P6.md.

The design is the
P4.1 spec
and its amendment (MI1 §8 without the S2 interception slice). The nested-m8i live acceptance passed on 2026-10-22.

Fixes

  • Supervisor crash orphaned syncs (#579, #558): a reconcile pass now picks up work orphaned when
    a supervisor crashes mid-flight.
  • Host capacity saturation (#597, #452): the vmpool sheds an idle standby before refusing, and a
    genuine refusal answers 503 instead of narrating the problem to the model.
  • sbx pull --apply refuses .codex/ (#594, #592) — Codex CLI's repo config, which can start MCP
    servers — on top of the existing .pi/ and .agents/ refusals.
  • sbx up warns about Claude Code's trust prompt and doctor names the recovery (#591, #577);
    sandbox agent steering fixed (#572, #568).
  • Context service retries the create-rollback cleanup instead of failing the roll-forward (#603).
  • The extension takes its two transport tiers separately (#606, #245), so the fast tier is not
    starved behind the slow one.
  • setup.sh propagates SH_MODEL_CATALOG_HOSTS into moca-settings (#575); the harness URL can be
    overridden (#589).
  • Dependency bumps for CVE-2026-111181 (body-parser) and four fast-uri CVEs.

Also in this release

  • Phase 2 of /workloads: per-workload sandboxes via a separate moca-provisioner, each
    workload's tasks sharing a PVC owned by MOCA (#588, #600), and a two-user Kind demo (#604).
  • The relay forwards trusted workload labels (#560); the delegation skill packs into
    .moca-delegate/ (#605); phase-2 design docs and the sbx release-channel notes.

Installing

curl -fsSL https://raw.githubusercontent.com/rossoctl/moca/main/scripts/install-mocactl.sh | sh

Installs the latest release (this one). The mocactl-edge prerelease tracks every push to main.

Full Changelog: v0.7.0...v0.8.0