name: workspace-sandboxes
description: "AgentScope local and sandboxed workspace backend workflows."
disable-model-invocation: true
metadata:
disco-role: operating
license: Apache 2.0
workspace-sandboxes
Use this sub-skill for workspace lifecycle, tool exposure, skill seeding, MCP seeding, archive import, and sandbox backend selection.
Read first
references/backend-matrix.md for the supported workspace backends and when to use each one.
references/lifecycle.md for initialize / list / add / remove / close workflows.
references/troubleshooting.md for backend, archive, skill, and MCP failures.
scripts/local_workspace_smoke.py for a safe local sanity check.
Typical triggers
- "How do I create a workspace?"
- "Which backend do I need for Docker, Bubblewrap, E2B, Daytona, K8s, or OpenSandbox?"
- "How do I seed skills or MCPs into a workspace?"
- "Why did archive import, path handling, or tool exposure fail?"
What belongs here
WorkspaceBase, LocalWorkspace, and sandboxed workspace backends
initialize, close, reset, list_tools, list_skills, list_mcps
add_skill, add_skill_archive, remove_skill, add_mcp, remove_mcp
- backend selection, archive import, and path-safety diagnostics
What does not belong here
- agent/tool/permission basics →
agent-core
- provider credentials and model families →
provider-connectors
- retrieval and memory workflows →
rag-memory
- service bootstrap and deployment →
service-platform
Use pattern
- Decide whether the workspace must be local, containerized, cloud-sandboxed, or Kubernetes-backed.
- Read the backend matrix before choosing the runtime.
- Use the local smoke script to confirm skill loading and file-tool exposure.
- Add MCPs and skills with the lifecycle methods, not by mutating the workspace directory manually.
- Read the troubleshooting page before changing backend-specific code.
Cross-links
- If the issue is really about the tool layer, use
agent-core.
- If the issue is about RAG or memory files, use
rag-memory.
- If the issue is a deployment topology problem, use
service-platform.