The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill instructs the agent to use specific command-line flags that bypass built-in security controls and human-in-the-loop permission prompts in sub-agents.
Recommends claude --permission-mode bypassPermissions for Claude Code, which allows the tool to execute actions without requesting user approval.
Recommends codex --yolo for the Codex CLI to enable autonomous command execution without verification.
[COMMAND_EXECUTION]: The skill provides patterns for executing arbitrary commands on remote and local systems through orchestration tools.
Instructs the agent to SSH into remote VPS environments and launch autonomous agents locally on those boxes.
Details the use of cmux send to relay strings directly to terminal surfaces, which can be used to execute shell commands.
[INDIRECT_PROMPT_INJECTION]: As an orchestration tool, the skill is designed to relay untrusted user data to other autonomous agents, creating a significant injection surface.
Ingestion points: Processes user-provided "work items" or "requests" and relays them to sub-agents (Codex, Pi, Claude, Hermes).
Boundary markers: Recommends wrapping prompts in plain double quotes but lacks robust delimiter instructions to prevent the sub-agent from misinterpreting the relayed data as instructions.
Capability inventory: The orchestrator can execute shell commands, manage files, and perform network operations via the sub-agents it controls (specifically mentioning cmux, ssh, and claude).
Sanitization: The skill lacks content sanitization; it explicitly advises against escaping characters (like quotes) for shell compatibility reasons, which may inadvertently allow shell injection if the relayed content is malicious.
[METADATA_POISONING]: The skill's YAML frontmatter includes a risk: critical field. While this may be a legitimate warning by the author, metadata fields are used to influence agent perception of its own safety and capabilities.

delegating-to-agents