The Agent Skills Directory

[PROMPT_INJECTION]: The skill handles 'delegation packets' which are untrusted data strings passed to worker agents, creating a surface for indirect prompt injection.
Ingestion points: External data enters via the <packet> argument in rtk herdr agent start commands or via follow-up prompts in rtk herdr pane run (documented in SKILL.md and references/plan-mode.md).
Boundary markers: The skill lacks technical boundary markers or sanitization logic for packet content, relying instead on high-level procedural instructions for the controller agent.
Capability inventory: The worker agents (claude, codex) have extensive capabilities, including filesystem access and command execution, which are invoked via the rtk orchestration tool.
Sanitization: The instructions mandate manual verification steps, including reviewing diffs and re-running tests, to mitigate the risk of malicious worker output.
[COMMAND_EXECUTION]: The skill executes shell commands to manage the orchestration environment, specifically using flags that disable safety prompts.
Evidence: Instructions in SKILL.md mandate the use of --dangerously-skip-permissions for Claude and --yolo for Codex during rtk herdr agent start to ensure workers run unattended without stalling on permission prompts.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for installing worker integrations from well-known providers.
Evidence: SKILL.md includes commands such as rtk herdr integration install claude and rtk herdr integration install codex to set up worker agents.

herdr-orchestration