CubeSandbox changelog: what's new each month
Every stable CubeSandbox release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 8 months covered; the current month updates daily.
November 2026 · so far
1 release: 0.8.0
November brought significant improvements to storage efficiency, scheduling flexibility, and observability. The month focused on making CubeSandbox more scalable and easier to operate in production environments.
Highlights
- New JuiceFS COW storage backend allows sharing templates and resuming sandboxes across different nodes.
- FS-only snapshots reduce storage costs by capturing only filesystem changes.
- Sandbox restart policies automatically recover from service crashes.
- Customizable scheduler plugin system with CEL expressions and external plugins.
- Sandbox metrics and logs now integrate with Prometheus and Grafana.
New
- JuiceFS COW storage backend (preview) for cross-node template sharing and snapshot booting.
- FS-only snapshot capability to reduce storage overhead.
- Automatic sandbox restart policies when internal services crash.
- Scheduler plugin system supporting CEL expressions and external gRPC plugins.
- Sandbox metrics collection for Prometheus and Grafana integration.
- Scheduling quality evaluation framework to test policies against simulated cluster topologies.
Improved
- E2B SDK compatibility extended to version 2.51.0 and later.
- Scheduler customization now more flexible with multiple plugin options.
October 2026
2 releases: kernel-release-261012-1 → 0.7.3
October brought stability improvements and performance enhancements focused on network reliability and faster sandbox restoration. The release includes a dedicated kernel assets package and significant optimizations to S3lvol operations.
Highlights
- Fixed intermittent network failures that affected sandbox access, improving overall network stability.
- S3lvol hot upgrade now reduces I/O pause time from 40 seconds to 8 seconds during component upgrades.
- Cross-node restore uses lazy-loading instead of full disk copying, making restores faster and more efficient.
- Gateway implementation is now customizable; users can replace it with any implementation following the contract.
- Template creation now validates nodes upfront and fails immediately with clear feedback when problems are detected.
New
- MCP URL and token support in Python, Node, and Go SDKs.
- Customizable gateway implementation with provided reference example.
Improved
- S3lvol cache performance for faster cross-node restore operations.
- S3lvol hot upgrade with dramatically reduced I/O pause during component upgrades.
- Cross-node restore now uses lazy-loading with local cache acceleration instead of full disk copying.
- Template creation experience with immediate upfront node validity checking.
Fixed
- Intermittent network failures when accessing sandboxes.
September 2026
3 releases: 0.7.1 → 0.7.2
September brought significant scalability improvements and snapshot enhancements to CubeSandbox. The month focused on faster restore operations, better multi-replica deployment, and more flexible storage options.
Highlights
- Standalone CubeTemplateCenter service enables independent scaling of template operations away from CubeMaster.
- Snapshot restore now completes in under 1 second with default MinIO setup, using optimized server-side copy and local caching.
- Snapshot support extended to Host Mount and Volume Plugin sandboxes with snapshot, restore, rollback, and clone operations.
- Active/standby disaster recovery now supported in lifecycle management service.
- Running sandboxes no longer block snapshot deletion, with storage reclaimed automatically after references are released.
New
- Whole-object local cache keeps immutable S3 data on local nodes to speed up repeated reads and re-imports.
- Local-filesystem storage option for component repositories and template artifacts alongside S3, PVC, and NFS.
- Template install path configuration in Cubelet allows large templates to be moved to the data disk.
- Standalone CubeTemplateCenter service with multi-replica deployment capability.
Improved
- Cross-node snapshot restore uses object-storage server-side copy for significantly faster transfers.
- Reduced read amplification during restore by coalescing scattered small reads into whole-object GETs.
- Concurrent fetch budget capped during restore to improve performance.
- Faster virtiofs operations on pause and snapshot paths through optimized state preparation flow.
- S3lvol hot upgrade no longer causes EIO errors inside sandboxes.
- virtiofsd hardening rejects root-directory mutations and forged readdirplus requests after directory whitelist is set.
Fixed
- Data loss in incremental snapshots.
- Running sandboxes blocking snapshot deletion.
- EIO errors during CubeS3lvol storage component upgrades.
August 2026
2 releases: kernel-release-260812-1 → 0.7.0
August brought significant architectural improvements and new storage capabilities to CubeSandbox. Version 0.7.0 streamlined the system by merging components and added support for remote snapshot storage and private image registries.
Highlights
- Templates and snapshots now track component versions, keeping historical snapshots usable after component upgrades.
- Network policy provisioning is faster, with sandbox creation latency significantly reduced for sandboxes with network rules.
- S3-based snapshot storage enables cross-node pause, resume, and snapshot cloning alongside the existing local XFS engine.
- Pull images from private HTTP registries when building templates.
- Remove nodes from the cluster to stop them participating in scheduling.
New
- S3 object storage backend for CubeS3lvol COW snapshots, enabling cross-node operations.
- Private HTTP image registry support for template building.
- MinIO deployed by default as the S3 volume backend.
- Node deletion capability to remove nodes from cluster scheduling.
- Version tracking in templates and snapshots for component dependencies.
Improved
- Merged NetworkAgent into Cubelet to reduce inter-process RPC calls during sandbox creation.
- eBPF network policy provisioning path optimized to reduce sandbox creation latency.
- Removed unnecessary netlink dump operations.
- Fewer components to deploy in the overall architecture.
- TAP device lifecycle management improvements.
July 2026
3 releases: 0.5.0 → 0.6.0
July brought resource optimization through auto-pause and auto-resume mechanisms, expanded ARM64 support, and improved lifecycle management across the control plane. The month focused on making sandboxes more efficient while maintaining compatibility and adding better operational tooling.
- SDK CreateOptions.Timeout is now an optional pointer; code relying on implicit defaults must be updated.
- EndAt timeout field now follows three-value semantics (-1 for no deadline) instead of previous behavior.
Highlights
- AutoPause and AutoResume automatically pause idle sandboxes and resume them on demand, freeing up resources.
- ARM64 support completed with MMIO-based hypervisor signaling and per-architecture kernel configuration.
- Sandboxes can now require traffic access tokens, adding network-level access control for private deployments.
- Control plane lifecycle management integrated with Redis service discovery for multi-replica deployments.
- SDK timeout handling redesigned to use optional values, eliminating hard-coded defaults.
New
- AutoPause mechanism tracks sandbox activity and pauses idle sandboxes based on configurable thresholds.
- AutoResume automatically resumes paused sandboxes when new requests arrive.
- Traffic access tokens enforce per-sandbox network access control for sandboxes with public traffic disabled.
- Idle timeout kill path with new POST APIs to manage sandbox lifecycle and timeout state.
- Cube Lifecycle Manager (CLM) coordinates pause and resume actions across multi-replica deployments.
- ivshmem device support in templates for shared memory operations.
- Short sandbox ID prefix resolution for easier operational queries.
- New cubemastercli volume-related subcommands for volume management.
Improved
- Configurable resource release ratio allows paused sandboxes to return a portion of their CPU and memory quota.
- CubeProxy now enforces traffic access tokens on all inbound requests.
- RPC deadline decoupled from idle TTL with separate create timeout parameter.
- SDK timeout fields (Go and Python) changed to optional pointers to avoid injecting defaults.
- Go SDK PTY capabilities improved for interactive terminal use.
- Python SDK X-API-Key header now carried on all request paths.
- Go SDK process handling and dataplane frame parsing improved.
- cubemastercli UX improvements for template and resource operations.
Fixed
- Go SDK exit-0 end event parsing corrected.
- Go SDK encode deduplication defect resolved.
- Dataplane defects in protocol frame handling fixed.
- CLM state sync ensures stronger sandbox lifecycle consistency between control and data planes.
- TencentCloud Terraform deployments now properly include cube-lifecycle-manager.
June 2026
3 releases: 0.3.0 → 0.4.0
June brought major snapshot and restore capabilities, networking and security improvements, and a new digital assistant console built on CubeSandbox. The month featured three releases focused on stability, performance, and expanding the platform's egress control and audit features.
- glibc >= 2.34 now required; installation fails on older distributions like CentOS 7 and Ubuntu 20.04.
- CUBESANDBOXNETWORKCIDR must be explicitly configured to avoid network conflicts.
Highlights
- Copy-on-Write snapshot engine using reflink volumes for efficient block-level snapshot operations.
- Incremental memory snapshots using soft-dirty page tracking dramatically reduce snapshot time and storage.
- Credential injection and domain filtering enable fine-grained egress policy control without exposing secrets to user code.
- AgentHub digital assistant console with full lifecycle management and snapshot/rollback API integration.
- Custom sandbox network CIDR configuration avoids conflicts with existing host subnets.
New
- CubeCoW Copy-on-Write snapshot engine with reflink-based volume snapshots.
- Per-cycle incremental memory snapshots via soft-dirty page tracking.
- Snapshot restore with proper VSOCK connection reset.
- AgentHub digital assistant console with persistence layer.
- Credential injection at proxy layer via EgressRule.inject.
- Domain filtering with policy-driven allow/deny lists for sandbox destinations.
- Access auditing with structured JSON logging of egress requests.
- Log streaming with stdout/stderr pipes for container logging.
Improved
- Snapshot write latency reduced by removing unnecessary syncall() calls.
- TAP TX offload enables checksum/TSO offload for redirected packets.
- Snapshot restore handles paused host-mount filesystems correctly.
- CubeVS fast-path hardened to reject SYN-only packets in port-mapping BPF.
- glibc preflight check fails fast on unsupported distributions.
- PVM consistency pre-check verifies CUBEPVMENABLE setting when kvmpvm module is loaded.
- Network-agent fix prevents TAP fd acquisition errors on restart.
- AgentHub preview feature added to WebUI with template store defaults.
Fixed
- Snapshot restore for sandboxes with paused host-mount filesystems.
- Network-agent TAP fd acquisition on restart causing existing sandbox network drops.
- Systemd 255 ExecStart/ExecStop errors by prefixing directives with /usr/bin/bash.
- Guest-initiated connection attempts bypassing egress policy in port-mapping BPF.
- CubeEgress version reporting now participates in component version matrix.
May 2026
3 releases: 0.2.0 → 0.2.2
May brought major usability improvements to CubeSandbox, including a new web management console and PVM deployment mode that removes bare-metal requirements. The month also addressed several critical bugs affecting disk QoS, host mounts, and network performance.
- Default sandbox exposed port changed from 8080/32000 to 49983 for E2B compatibility.
- CubeMaster database migration required for new requestid column on template-image jobs.
Highlights
- Web Management Console now provides visual cluster overview, node status, template management, and API key administration.
- PVM Deployment Mode enables CubeSandbox to run on ordinary cloud servers without bare-metal or nested virtualization.
- Python SDK reached feature parity with full sandbox lifecycle, code execution, and network policy support.
- Fixed disk QoS limits that were silently ignored due to annotation key mismatch.
- Resolved network performance issue caused by advertised hardware offload features creating malformed packets.
New
- Web Management Console with cluster overview, node and sandbox status, template management, and API key administration.
- PVM Deployment Mode powered by Pagetable-based Virtual Machine for ordinary cloud servers.
- Custom DNS flag (--dns) in cubemastercli template for specifying DNS server during template creation.
- Direct-connect transport in Python SDK honoring CUBEPROXYNODEIP via TCP direct connection.
- Network policy as first-class parameters with allowinternetaccess and explicit allow/deny lists.
- Full sandbox lifecycle in Python SDK including create, connect, pause, kill, list, and health operations.
- Code execution via runcode() with streaming stdout/stderr, environment variables, and timeout support.
- 12 worked examples and benchmark.py for concurrent create/delete performance testing in Python SDK.
Improved
- Host mount propagation by switching Cubelet mount namespace from private to slave mode.
- Cubelet kernel file handling by splitting verification into separate fast-path and refresh operations, skipping SHA256 on startup.
- CubeAPI web endpoints added for management console integration.
- Python SDK now uses single HTTP client (httpx only) with typed Config and deterministic cleanup.
- Template-image job creation now idempotent via requestid column with unique index.
- E2B compatibility by changing default exposed port from 8080/32000 to 49983.
- Command timeout configurability in Cubelet via storage plugin TOML config, replacing hardcoded 3-second default.
- Artifact runtime file materialization for PVM templates simplified and fixed.
Fixed
- Disk QoS limits silently ignored due to Cubelet reading annotation with wrong key.
- Host-mount requests silently dropped due to key mismatch between CubeAPI and CubeMaster.
- DeadGC permanently freezing paused sandboxes by avoiding mutex contention during state calls.
- Virtio-net TAP offloads causing network corruption by disabling TSO/UFO/CSUM in hypervisor.
- Cubecli exec nil-pointer dereference panic on stdin EOF.
- Duplicate template-image job creation via request idempotency.
April 2026
3 releases: 0.1.0 → 0.1.2
April brought the initial release of CubeSandbox with core isolation and performance features, followed by development environment improvements and deployment optimizations.
Highlights
- Blazing-fast cold start through resource pool pre-provisioning
- True kernel-level isolation with dedicated kernel per Agent
- Zero-cost migration as a drop-in E2B replacement
- Network security enforced via CubeVS with eBPF
- High-density deployment on single nodes with memory reuse
New
- CubeSandbox base image and entrypoint script
- QEMU-based development environment for OpenCloudOS 9
- China mainland CDN mirror support for faster downloads
- CubeProxy with optimized package mirror configuration
Improved
- CubeProxy switched to tencentyun mirror for better performance
- Integration tests enabled for hypervisor
- Rustup mirror made configurable via build arguments
- Development environment documentation updated
Fixed
- CubeAPI now maps missing sandbox template to bad request on create
- Certificate path documentation corrected from shell command to absolute path
- CI workflow improved with GHREPO environment variable support
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. CubeSandbox: pricing, features and alternatives · All changelogs