The Agent Skills Directory

[COMMAND_EXECUTION]: The execution template in SKILL.md (bash {baseDir}/scripts/config_reader.sh FIELD DEFAULT) is vulnerable to shell command injection. The placeholders FIELD and DEFAULT are replaced with values derived from the agent's $ARGUMENTS. If these arguments contain shell metacharacters like ;, &, or |, an attacker could execute arbitrary commands in the host environment.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from local configuration files which could be manipulated to influence agent behavior.\n
Ingestion points: .agents.yml and .agents.local.yml files located in the project root (accessed via scripts/config_reader.sh).\n
Boundary markers: None. Values are extracted and returned as raw strings to the agent context.\n
Capability inventory: The skill reads file content and provides it to the agent, which uses these values for configuration settings like tech_stack or workflow.\n
Sanitization: None. The script does not perform any validation or sanitization of the values read from the YAML files before returning them.\n- [EXTERNAL_DOWNLOADS]: The skill requires the yq utility to function. Both the SKILL.md and scripts/config_reader.sh files provide instructions for downloading and installing yq from its official GitHub repository or via reputable package managers such as Homebrew and Snap. These references target a well-known and established open-source tool.

config-reader