server/v0.1.12
opensandbox-group/OpenSandboxserver/v0.1.12Apr 25, 2026by Pangjiping
AI Summary
Introduces Windows Sandbox support and GPU resource limits, while changing host path defaults to deny-all for security.
Key Highlights
- Windows Sandbox support and BatchSandbox Kubernetes workloads
- GPU resource limits honored across Docker and Kubernetes runtimes
- X-Forwarded-* headers injection for proxied HTTP requests
- allowed_host_paths defaults to deny-all for security
Breaking Changes
- allowed_host_paths defaults to deny-all
- Server requires explicit startup confirmation when api_key is unset
New Features
- Windows Sandbox support
- GPU resource limits (Docker and Kubernetes)
- Proxy header injection
- EIP support for proxied endpoint URLs
Full Release Notes
## What's New ### ⚠️ Breaking Changes - **`allowed_host_paths` defaults to deny-all** — Previously, when `[storage].allowed_host_paths` was empty or unset, the server allowed all host path mounts into sandboxes, creating a sandbox escape vulnerability (closes #750). This release flips the default to deny all host mounts. Users who need the legacy behavior must explicitly set `allowed_host_paths = ["/"]` as a temporary compatibility measure. Additionally, the server now requires explicit startup confirmation when `api_key` is unset. (#751) ### ✨ Features - **Windows Sandbox** - windows server by dockur/windows (#712) - resolves a port exposure logic error affecting Windows sandbox containers and egress network scenarios. (#733) - Windows container support for BatchSandbox Kubernetes workloads. (#752) - propagate opensandbox.extensions. to Pod annotations — Extensions with the `opensandbox.extensions.` prefix are now automatically copied to Kubernetes Pod annotations with the `opensandbox.io/extensions.` (#772) - **GPU resource limits now honored across both runtimes** - **Docker**: mapped to `DeviceRequest(count=N, capabilities=[["gpu"]])`. (#775) - **Kubernetes**: translated to the canonical extended resource `nvidia.com/gpu`, enabling the NVIDIA device plugin to schedule GPU pods. (#782) ### 🐛 Bug Fixes - Inject X-Forwarded-* headers for proxied HTTP requests — When proxying HTTP requests to a user sandbox via `use_server_proxy`, the server now injects `X-Forwarded-Proto`, `X-Forwarded-Host`, and `X-Forwarded-For` headers. Previously, web apps inside the sandbox (VS Code, Jupyter, OAuth callbacks) had no way to determine the original scheme, host, or client IP, causing broken HTTPS redirects and incorrect absolute URL generation. Uses `setdefault` semantics for upstream proxy chain safety. (#777) - Honor `server.eip` for proxied endpoint URLs — When `use_server_proxy=true`, the server now uses the configured `server.eip` to generate externally reachable proxy endpoints instead of returning internal `base_url` addresses. Falls back to existing behavior when `eip` is unset. (#747) ### 📦 Misc - bump `execd` to v1.0.13 with config template, documentation, and test updates (#763) - chore(deps): bump python-dotenv from 1.2.1 to 1.2.2 in /server (#784) ## 👥 Contributors Thanks to these contributors ❤️ - @bytkim - @wangdengshan - @504097723 - @Pangjiping - @ninan-nn --- - PyPI: opensandbox-server==0.1.12 - Docker Hub: opensandbox/server:v0.1.12 - Aliyun Registry: sandbox-registry.cn-zhangjiakou.cr.aliyuncs.com/opensandbox/server:v0.1.12