The Agent Skills Directory

[COMMAND_EXECUTION]: The skill manages agent execution through extensive use of shell commands, including git worktree management, tmux session control, and native agent management CLIs as seen in backends.md and the native recipe files.
[PRIVILEGE_ESCALATION]: As described in backends.md and native recipes, the skill instructs the agent to use sensitive bypass flags like --dangerously-skip-permissions for Claude Code and --dangerously-bypass-approvals-and-sandbox for Codex. These flags allow sub-agents to operate without interactive human approval.
[DYNAMIC_EXECUTION]: The skill employs dynamic shell execution, specifically using eval to integrate the output of scripts/detect-backend.sh into the agent's environment. It also manages the lifecycle of dynamically spawned sub-agents.
[INDIRECT_PROMPT_INJECTION]: The skill functions as a dispatcher for sub-agent prompts, creating a surface for indirect prompt injection. (1) Ingestion points: Untrusted prompts enter through the spawn() and nudge() verbs in SKILL.md. (2) Boundary markers: The check() operation in backends.md uses isolation, but other spawn paths rely on soft prompt contracts. (3) Capability inventory: Sub-agents can perform file operations, subprocess execution, and further agent spawning as detailed in the native backend recipes. (4) Sanitization: The skill does not perform automated sanitization, relying instead on model instructions and prompt boundaries.

aep-executor