The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using the Bash tool that incorporate potentially unvalidated user input. Specifically, the goal, target, and slug variables are directly interpolated into command strings for gh copilot suggest, gemini, and internal scripts such as run_agent.py and experiment_log.py. This creates a significant risk of command injection if a maliciously crafted goal or target name is provided.
[EXTERNAL_DOWNLOADS]: The skill utilizes well-known external CLI services, specifically GitHub Copilot (gh) and Google Gemini (gemini), to perform brainstorming tasks. These services involve network communication and the processing of remote model outputs.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection.
Ingestion points: Target files (e.g., SKILL.md) are read via the Read tool in Phase 2. User-provided Evolution goal and Target are also ingested as primary inputs.
Boundary markers: None identified; untrusted data is directly interpolated into prompts and shell commands without delimiters or explicit instructions to ignore embedded commands.
Capability inventory: The skill has access to Bash, Read, and Write tools, and executes local Python scripts and external CLIs.
Sanitization: No sanitization, escaping, or validation of external content or user input is present before interpolation into prompts or shell commands.

os-evolution-planner