The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes gh (GitHub CLI), endorctl (Endor Labs CLI), and jq for read-only data retrieval and processing. It includes specific safeguards, such as using set -o pipefail and separating stderr from JSON projection, to ensure reliable and safe command execution.
[EXTERNAL_DOWNLOADS]: The agent interacts with GitHub.com and Endor Labs APIs to collect repository and project metadata. These operations are restricted to authorized vendor endpoints and are central to the skill's intended functionality.
[PROMPT_INJECTION]: To protect against indirect prompt injection from untrusted repository content, the skill explicitly instructs the agent to treat all repository files, dependency metadata, and command outputs as data rather than instructions. This clear separation of data and control helps prevent the agent from being manipulated by content within analyzed repositories.
[CREDENTIALS_UNSAFE]: The skill implements best practices for secret management by forbidding the agent from reading entire configuration files or echoing sensitive tokens. It limits local configuration access to the extraction of the specific namespace key only.

probe-droid