The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs local command execution of docker compose config to synchronize project runtime settings. It explicitly utilizes jq to filter the output, ensuring that sensitive data such as environment variables and secrets are not ingested into the agent's context.
[DATA_EXPOSURE_&_EXFILTRATION]: The skill includes defensive instructions to prevent the reading of global agent configuration files or personal memory, restricting its scope to the active workspace to preserve privacy.
[REMOTE_CODE_EXECUTION]: While the skill references external tools (e.g., Claude CLI, Cursor CLI) and other 'companion' skills, it does not perform automated remote downloads or execution of untrusted scripts. It treats these as pre-existing environmental tools.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from *.code-workspace and Docker Compose files. It implements boundary markers (version markers) and sanitization (via jq) to mitigate the risk of processing malicious instructions embedded in these files.

agents-md