The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill is designed to automatically download and install code from external GitHub repositories using the bin/add-skill command. Even though auto-installation is restricted to sources defined in a local trusted-sources.txt file, this mechanism creates a direct pathway for remote code to be integrated and executed within the agent's environment.
[DATA_EXFILTRATION]: Search queries are derived from internal repository data, including log snippets from memory/logs/, issue titles from memory/issues/INDEX.md, and failure reports from memory/cron-state.json. These queries are transmitted to external services such as the npx registry and the skills.sh search engine, which may inadvertently expose internal operational details or sensitive gap information.
[PROMPT_INJECTION]: During the search and evaluation process, the skill ingests and processes descriptions from external, untrusted skills. This behavior creates a surface for indirect prompt injection, where malicious instructions embedded in a remote skill's metadata could influence the agent's scoring logic or PR generation process.
[COMMAND_EXECUTION]: The skill executes various shell commands to manage its lifecycle, including npx to find external packages and gh api to check repository status. These operations involve dynamic interaction with external registries and APIs based on strings derived from potentially untrusted search results.

search-skill