The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The orchestrator agent is designed to ingest and integrate reports generated by sub-agents. This creates a vulnerability surface where a sub-agent (or the model it uses) could provide a report containing hidden instructions to manipulate the orchestrator.\n
Ingestion points: The command collect (implemented in scripts/lib/commands/collect.mjs) reads report files from the .herdr-agents/reports/ directory.\n
Boundary markers: Reports are Markdown files, but there are no specific markers or instructions for the orchestrator to ignore executable content within them when integrating their findings.\n
Capability inventory: The orchestrator agent possesses extensive capabilities, including executing shell commands, spawning new agents, and modifying repository files via its primary orchestration role.\n
Sanitization: While the skill includes utility functions to redact secrets from its own logs (scripts/lib/text.mjs), it does not sanitize or filter the natural language content of sub-agent reports before they are processed by the orchestrator.\n- [COMMAND_EXECUTION]: The skill's primary function involves executing various CLI tools (e.g., herdr, claude, codex, grok, cursor). Although it uses spawnSync without a shell to mitigate command injection risks, it provides the AI agent with a framework to manage and interact with multiple concurrent processes, which is a powerful capability that requires careful oversight.

herdr-agents