opensandbox-group/OpenSandbox

Why do you need it? DeepSeek Harness (dsh) provides a plugin-based agent harness. We would like to build an OpenSandbox plugin that connects it to isolated execution environments powered by the fast-sandbox runtime. The goal is to combine dsh’s agent capabilities with OpenSandbox’s sandbox lifecycle and execution APIs, using fast-sandbox for fast provisioning and efficient resource usage. This could become the foundation for a self-hosted product similar to Managed Agents: users submit tasks, agents work in dedicated environments, and users can follow progress, retrieve artifacts, and return to existing sessions. We are opening this issue to find contributors interested in designing and building the integration. How could it be? The proposed architecture is: DeepSeek Harness │ OpenSandbox plugin │ OpenSandbox API / SDK │ fast-sandbox runtime │ Isolated execution environments The plugin should consume OpenSandbox’s public APIs. Runtime provisio

Summary Kubernetes Pool E2E (ingress-header) intermittently fails in tests/python/tests/test_sandbox_pool_e2e_sync.py::TestSandboxPoolSingleNodeE2ESync::test_concurrent_start_shutdown_stress_single_node: E AssertionError: Timed out waiting for pool remains usable after lifecycle stress FAILED tests/test_sandbox_pool_e2e_sync.py::TestSandboxPoolSingleNodeE2ESync::test_concurrent_start_shutdown_stress_single_node Where it happened Workflow: Kubernetes Mini E2E CI → job Kubernetes Pool E2E (ingress-header) → step Run Kubernetes Pool E2E Run: https://github.com/opensandbox-group/OpenSandbox/actions/runs/36513861015/job/109231868754 Surrounding log noise (may or may not be related) WARNING opensandbox.sync.pool - Pool sandbox cleanup failed: ... error=Kill sandbox ... not found | [KUBERNETES::SANDBOX_NOT_FOUND] WARNING opensandbox.sync.pool - Acquire renew failed after idle connect; ... Cannot acquire from a retired pool run ... state=DRAINING ERROR asyncio - Task was destroyed but

Summary #1497 is still reproducible on main. #1523 attempted a fix but has been closed: it drifted ~535 commits behind main, was in a conflicting state, and its CI never ran beyond auto-label, so the 218 lines of new tests were never executed. This issue tracks a clean rewrite against current main, preserving the parts of #1523 that were reviewed and validated. The defect on main today kubernetes/internal/controller/batchsandbox_status.go:362, inside applySteadyRuntimePhase: if status.Phase == sandboxv1alpha1.BatchSandboxPhaseFailed { return } Once a BatchSandbox reaches Failed, the steady-state path returns unconditionally. If the same Pod (same UID) later recovers to Running and Ready, the phase is never recomputed and the BatchSandbox stays Failed forever. This is exactly the scenario reported in #1497. Prior art from #1523 worth keeping These were reviewed and confirmed correct; they should carry over rather than be redesigned: Recovery keyed on Pod UID, not name. A replacemen

Background The fast-sandbox integration test environment (scripts/fast-sandbox-env/integration-env.sh) currently uses MinIO (minio/minio + minio/mc) as the S3-compatible artifact store for SandboxTemplate builds/publish and snapshot manifests. We want to replace MinIO with RustFS — an S3-compatible, open-source (Apache-2.0) object storage written in Rust — as the artifact store for this environment. Scope Primary target: scripts/fast-sandbox-env/ (the up / status / down stages): Replace the MinIO container with RustFS in integration-env.sh (minio_up stage): container image, run flags, data volume, credentials env health check — the current probe uses MinIO-specific /minio/health/live bucket bootstrap — currently mc alias set + mc mb Replace minio/mc client usages (mc alias, mc mb, mc stat for manifest size checks) with a RustFS-compatible approach (RustFS tooling or generic S3 client) Update env vars / toggles: MINIO_IMAGE, MINIO_PORT, MINIO_CONSOLE_PORT, MINIO_AK, MINIO

Background The e2e test suite contains a number of tests targeting the code-interpreter API. These tests depend on the code-interpreter SDK and a dedicated image, and carry a high maintenance cost. They should be removed across all languages. Test files to remove Language File Python tests/python/tests/test_code_interpreter_e2e.py Python tests/python/tests/test_code_interpreter_e2e_sync.py Go tests/go/code_interpreter_e2e_test.go JavaScript tests/javascript/tests/test_code_interpreter_e2e.test.ts C# tests/csharp/OpenSandbox.E2ETests/CodeInterpreterE2ETests.cs Related cleanup .github/workflows/real-e2e.yml: remove the run_code_interpreter_e2e input and the RUN_CODE_INTERPRETER_E2E env vars from all jobs tests/csharp/OpenSandbox.E2ETests/OpenSandbox.E2ETests.csproj: remove the ProjectReference to sdks/code-interpreter/csharp (after confirming no other tests use it) tests/javascript/package.json: remove the @alibaba-group/opensandbox-code-interpreter dependency and u