The Agent Skills Directory

[COMMAND_EXECUTION]: The skill manages the execution of setup commands (e.g., pnpm install) and agent processes within sandboxes.
[EXTERNAL_DOWNLOADS]: Provisioning involves cloning Git repositories and downloading container images (e.g., node:22) to establish the workspace.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external code and instructions into the sandbox environment. 1. Ingestion points: defineWorkspace source and instructions fields. 2. Boundary markers: instructions are written to AGENTS.md and symlinked as CLAUDE.md. 3. Capability inventory: SandboxHandle provides filesystem access, git operations, and process execution (exec/spawn). 4. Sanitization: No explicit sanitization of repository content is mentioned, as the framework relies on provider isolation (Docker, Daytona) to contain impact.
[PRIVILEGE_ESCALATION]: The documentation provides instructions for using sudo -n within specific sandbox providers (e.g., Daytona) to allow package installation during the setup phase.

ai-sandbox