The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's documentation includes examples for installing development tools, such as the Cursor CLI, using a remote script from its official domain.
[REMOTE_CODE_EXECUTION]: The SDK facilitates the execution of arbitrary shell commands through configuration fields like setup_commands, run_cmd, and custom_install_cmd to enable automated environment preparation and agent benchmarking inside sandboxes.
[PROMPT_INJECTION]: An indirect prompt injection surface exists because the skill is designed to ingest and process external benchmark datasets (such as SWE-bench) which may contain adversarial instructions.
Ingestion points: The DatasetConfig field in YAML configurations specifies external sources for benchmark data.
Boundary markers: The provided configuration templates do not include specific delimiters or warnings to prevent the agent from following instructions embedded within the datasets.
Capability inventory: The skill possesses extensive execution capabilities, including shell command execution and package installation within the evaluation environment.
Sanitization: No explicit sanitization or validation logic for external dataset content is described in the provided documentation.
[CREDENTIALS_UNSAFE]: The documentation guides users on managing sensitive credentials, such as API keys and access tokens, using environment variables and placeholders, which follows standard security practices for secret management.

rock-agent-sdk