The Agent Skills Directory

[COMMAND_EXECUTION]: The helper script scripts/corral_agents.py utilizes subprocess.run to orchestrate interactions with local system tools including git, herdr, and the native corral plugin. Commands are constructed as argument lists (arrays) rather than raw shell strings, which effectively mitigates common shell injection vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection as it ingests user-supplied task descriptions and prompts (via CLI arguments or JSON batch files) and forwards them to downstream agent processes.
Ingestion points: Arguments passed to scripts/corral_agents.py (--task, --prompt) and JSON objects processed by the batch command.
Boundary markers: The script does not wrap user prompts in explicit security delimiters, but it does combine them with predefined preset instructions.
Capability inventory: The script has the capability to execute arbitrary CLI agents and read/write files within the repository context.
Sanitization: The script performs validation on task titles (length limits and whitespace normalization) and includes logic to redact potential API keys and full prompt text from status summaries.
[DATA_EXPOSURE]: The skill performs discovery of local configuration files such as presets.toml and plugins.json to resolve paths and environment variables. To prevent accidental leakage of sensitive information, the script implements a redaction mechanism in its redacted_command and sanitized_preset functions to mask API keys and large prompt blocks before printing them to the console or JSON summaries.

corral-launch-agents