The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Bash script (scripts/delegate-sol.sh) which serves as an orchestrator for the herdr, jq, and codex command-line tools.
[COMMAND_EXECUTION]: The orchestration script spawns new subagents using the danger-full-access security profile and the -a never flag, which bypasses manual approval prompts. This grants the delegated agents unrestricted read/write access to the filesystem and the ability to execute commands autonomously within the Herdr environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-defined task descriptions and interpolates them into instructions for delegated subagents, creating a surface for indirect prompt injection.
Ingestion points: Task content is ingested through the --task argument, the --task-file path, or standard input in scripts/delegate-sol.sh.
Boundary markers: The script wraps the user task within a structured prompt template using clear headers (e.g., 'Task:', 'Rules:', 'Return exactly these sections:') to separate user input from system-level instructions.
Capability inventory: Delegated subagents possess full filesystem access and the ability to execute arbitrary shell commands via the herdr workspace.
Sanitization: The task text is stored in a temporary file with restricted permissions (chmod 600) to ensure only the intended subagent can access the full prompt content during initialization.

delegate-sol