The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents the local-command adapter, which is designed to execute arbitrary shell commands via sh -c. Examples provided in the documentation and examples/ directory demonstrate building, linting, and testing scripts that interact directly with the local environment.
[REMOTE_CODE_EXECUTION]: By providing the syntax and environment details for the local-command feature, the skill enables the execution of user-defined scripts on the host system within a temporary sandbox. This mechanism allows for dynamic code execution triggered by task completion events.
[EXTERNAL_DOWNLOADS]: The documentation includes examples (e.g., in harness-python.yaml and harness-rust.yaml) that use package managers such as npm, pip, and cargo to fetch and install dependencies from external registries at runtime. These operations target well-known services like the npm registry, PyPI, and Crates.io.
[DATA_EXFILTRATION]: Execution environments for the local-command adapter include the HARNESS_WITNESS_REPO variable, which provides access to the full source repository, including the .git history. This access, combined with the potential for network operations through package managers or custom commands, creates a surface for sensitive data exfiltration.
[INDIRECT_PROMPT_INJECTION]: The skill describes a workflow where configuration is ingested from the harness.yaml file to drive automated actions. This introduces an attack surface for indirect injection if the configuration file is modified by an untrusted party.
Ingestion points: The settings.gates section of the harness.yaml file.
Boundary markers: No specific delimiters or "ignore" instructions for embedded data are mentioned.
Capability inventory: The system supports arbitrary shell command execution via the local-command adapter.
Sanitization: The instructions do not specify any sanitization or escaping of the command strings before they are passed to the shell.

gate-customization