The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill requires agents to be launched using flags that explicitly bypass security protocols. This includes --dangerously-skip-permissions for Claude, --approve-for-me for Codex, and a combination of --yolo and --sandbox disabled for the Cursor Agent, which effectively removes execution constraints.- [COMMAND_EXECUTION]: The skill instructs the agent to execute multiple shell commands via the herdr CLI to manage terminal panes and spawn background processes with the aforementioned elevated privileges.- [INDIRECT_PROMPT_INJECTION]: The skill ingests arbitrary conversation context via a 'snapshot' and provides it to unrestricted sub-agents as an authoritative brief. This creates a vulnerability where untrusted data from the conversation can control the actions of a sub-agent operating without safety filters.
Ingestion points: Conversation snapshots captured in Step 1 of the dispatch workflow.
Boundary markers: Absent; the sub-agent is instructed to treat the snapshot-based prompt as authoritative even in the event of conflict.
Capability inventory: Full shell execution capabilities via herdr and agent-specific CLIs (claude, codex, agent, pi).
Sanitization: None; the snapshot is passed directly to the next agent without filtering.

handoff-that-to-agent-pane