Repository navigation
Description
Parent: #535. Phase 4 (runnable cold profile). Depends on the shared host foundation #538, guest root/mapping core #539 and streaming exec #537; real cold acceptance also requires source-bound VMM #547 (no dormant virtio-fs on zero-export direct boot) and #548 (verified opened objects and pinned grants). Their independent VMM work need not block host/guest foundation implementation. Coordinate CLI/EROFS-scratch/artifact surfaces with #359, but do not make its unfinished Python-parity CLI port a hard prerequisite. Supply the minimal standalone adapter/commands here and reuse #359's CLI when it lands; do not fork nvx.py or a second policy engine. #358 must supply an accepted Rust guest build before claiming source-bound end-to-end results.
Goal
Run the image-backed sandbox cold path entirely from nvx-rs and its Rust guest with a versioned NVXC application adapter. A caller provides an explicitly authorized local read-only image/layer set, runtime bundle, resources and policy. Another runtime can use the same host/guest mechanics through a different application transport. A remote image registry, product error/C ABI, and unrelated container orchestration do not belong in this standalone profile.
Required flow and ownership
- Validate input and an explicit local host-object/proxy allowlist. Pin immutable runtime/image identity at admission; reject exposing state/runtime/cache/template roots. Bundle verification occurs at selection/load, not by hashing large binaries on start/exec.
- Use nvx-rs: share image-backed sandbox host launch, policy, and VM supervisor primitives #538 to plan the exact
ExportPlanand network/proxy policy, emit a typed OpenVMM launch with a cold-plugged read-only image and openvmm: verify and receipt opened host attachments #548 opened-media receipt. For zero exports, select OpenVMM prerequisite: omit dormant virtio-fs on direct no-map cold boots #547's opt-in no-filesystem-device VMM topology; mapped cold launches attach an aggregate. A NIC is present only when connectivity is requested. Persist an owned VM record with PID+start-time/endpoint/capability and a recoverable launch-log claim. - Guest PID 1 brings up minimal platform/time/namespace facilities, then a restricted authenticated control service before assembling the image overlay. Accept a bounded/versioned configuration; use nvx-rs: factor minimal guest rootfs and descriptor-rooted host mapping #539 for the selected read-only lowers and tmpfs/ext4-scratch upper and safe ordered/zero mappings. Configure the guest interface, DNS and proxy only after the VMM policy is enforced. Ready echoes the accepted configuration identity; no exec before Ready.
- Implement
validate,provision,start,exec, bounded guest logs,stop, diagnostics anddeprovisionin the standalone adapter. nvx-rs: add bidirectional credit-based streaming exec to the matched host and guest #537 supplies one-active-exec/bounded-wait/credit behavior. This image-backed caller fixes root to a documented ceiling ofCAP_CHOWN,CAP_DAC_OVERRIDE,CAP_FOWNER,CAP_FSETID,CAP_KILL,CAP_NET_BIND_SERVICE,CAP_SETGID,CAP_SETUIDandCAP_AUDIT_WRITE, with other bounding/ambient capabilities dropped andno_new_privs; noCAP_SYS_ADMIN,CAP_MKNOD, raw-device or tracing powers. Private workload mount/PID namespaces and bounded cgroup resources protect PID 1. Raw exports, host control, configuration regions and device nodes remain inaccessible; nested read-only paths inside writable mappings must be host-enforced or demonstrably non-bypassable under this ceiling, otherwise rejected. Preserve existing non-root behavior of other public paths without a selectable isolation-profile flag. - Recover/reattach a verified VMM after a host process restart, restore console/log ownership, and fail closed on partial launch, changed artifacts or mismatched guest identity. Stop takes priority over the finite cross-process exec queue and cancels its waiters. Attempt graceful guest shutdown; if an active exec keeps the sole broker session, terminate only the verified VM after the bounded timeout rather than claiming guaranteed graceful delivery.
Acceptance
- A reproducible source-bound local bundle, including the static guest initramfs, minimal standalone commands and a reproducible public single-GPT image fixture (documenting p1 and p2+ roles), runs image-backed cold workloads end-to-end without ttrpc, protobuf, OCI/runc, a product image registry, or implicit unauthenticated host roots. An old running bundle never receives an unpinned new protocol after a host upgrade.
- Both tmpfs and ext4-scratch upper modes work with their respective read-only layer formats. A no-map/no-network, single-workload cold launch omits the virtio-fs, NIC and image slot from the actual VMM and guest inventories using OpenVMM prerequisite: omit dormant virtio-fs on direct no-map cold boots #547; no
--mountor guest mount is not enough. Mapped cold launches retain the aggregate slot; work for extra children/mappings occurs only when requested. - Measure cold-to-Ready and exec-start p50/p95, guest size and in-flight memory against source-matched artifacts and set a regression budget before declaring this profile efficient. Break out pre-ready control attach/activation latency until nvx-rs: replace activation RPC delivery with an uncaptured VMM configuration region #544 replaces its delivery; do not add per-start/exec large-file hashing or omit a required readiness step.
- Cold no-network launches omit the NIC; image and object-pinned mapped roots are verified against the actual opened VMM handles and receipts from openvmm: verify and receipt opened host attachments #548. A renamed pinned file child never exposes its replacement, while an explicitly live-path child keeps follow-name behavior; mapped paths and network/proxy policies match host enforcement and guest configuration.
- Ready/build identity, logs, shutdown, error outcomes and restart/reconciliation match the common lifecycle conformance matrix; credit streams handle arbitrarily long total output with bounded in-flight memory. Tests with separate CLI processes prove queue limits, priority Stop, verified force-stop fallback, root capability bounds, failed remount/user-namespace bypass, inaccessible raw devices/export and no surviving workload descendants.
- Targeted KVM/MSHV/WHP runs where supported record exact host, guest, VMM and artifact revisions; no skipped stage is reported as a pass. The existing managed profile and nvx-rs: make the nvx-rs CLI the backend of nvx.py sandbox, egress, and artifact lookup #359 CLI contract remain intact.
Activity
[-]nvx-rs: provide a standalone trusted-sandbox cold-start profile[/-][+]nvx-rs: provide a standalone image-backed sandbox cold-start path[/+]esaurez commentedon Oct 20, 2026
Replaced by the rewritten #535. The guest part of the cold-start path moves to #621 (A3), the host part to #622 (A4), and its retained acceptance to #624 (A6). Closing in favor of those issues.