The Agent Skills Directory

[COMMAND_EXECUTION]: The skill manages agent teams by executing various CLI tools including git for repository management, herdr for pane and agent control, vercel for deployment operations, and claude for relaying messages via one-shot sessions.
[DATA_EXPOSURE]: The skill reads agent session transcripts from ~/.claude/projects/ and ~/.codex/sessions/ using the scripts/ctx.sh script to monitor token usage and context window consumption.
[INDIRECT_PROMPT_INJECTION]: The orchestrator logic ingests reports and plans from multiple sub-agents, which may include untrusted or sandboxed sessions, creating a vulnerability to indirect prompt injection.
Ingestion points: Reads teammate reports from .agents/tmp/ files, repository-level instructions (AGENTS.md, CLAUDE.md), and cross-session messages.
Boundary markers: Mentions the use of unique markers for certain message relays but lacks a comprehensive strategy for delimiting all external input.
Capability inventory: Possesses significant capabilities including file system modification, subprocess execution of management tools, and deployment via external services.
Sanitization: Does not specify methods for sanitizing or validating the content of sub-agent reports before they are used to drive the orchestration logic.

orchestrator