The Agent Skills Directory

[COMMAND_EXECUTION]: The skill manages the execution of multiple external AI agent CLI tools including Claude Code, Codex, Cursor, Grok, and OpenCode. For 'writable' turns (implementation tasks), the instructions explicitly direct the use of bypass flags such as --permission-mode bypassPermissions, danger-full-access, --force, and --auto. These flags are used to allow unattended operation by removing interactive user approval prompts for actions like shell command execution.\n- [PRIVILEGE_ESCALATION]: By utilizing the aforementioned bypass permissions, the skill enables agents to perform actions that would typically require higher-level authorization or explicit user consent in their standard operating modes. This design is documented as a requirement for autonomous operation in dedicated developer environments but represents a significant escalation of the agent's effective capabilities.\n- [INDIRECT_PROMPT_INJECTION]: The skill has a defined surface for processing untrusted data from external sources.\n
Ingestion points: Transcripts are read via herdr pane read in scripts/herdr-pair.mjs and reply files are read from disk in scripts/pair-headless.mjs.\n
Boundary markers: Present. The protocol uses structured headers like [agent <from> -> <to> kind=<kind> sid=<sid>] to delimit agent communications.\n
Capability inventory: The skill executes agent CLI binaries with high autonomy using spawn and execFileSync in scripts/herdr-pair.mjs and scripts/pair-headless.mjs.\n
Sanitization: Present. The skill implements a shellQuote function to escape arguments before they are interpolated into shell commands.

pair