The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill manages the ingestion of untrusted user queries for live information (e.g., weather, market data). It mandates the use of strict boundary markers (such as 'freshness_boundary/v1' and 'prepared-vs-observed boundary') to distinguish between guidance and external data. The skill's capabilities are limited to status logging via a specific runtime tool ('omh'), and it lacks direct file-write or network-send functions, relying instead on external harnesses for actual data retrieval. It requires reporting facts only from observed provider evidence, which serves as a logical sanitization layer to prevent the agent from obeying instructions embedded in retrieved data.
[COMMAND_EXECUTION]: The skill includes instructions to use a specific command-line tool ('omh') for recording workflow status during execution. These commands are static, used for internal state management within the framework, and do not incorporate unvalidated user input or access sensitive file system locations like SSH keys or environment configuration files.
[OBFUSCATION]: The documentation includes routing keywords in multiple languages, including Korean Hangeul (e.g., '오늘 날씨'). These are standard identifiers used for natural language processing to improve the agent's routing accuracy across different locales and do not represent attempts to hide malicious intent through homoglyphs or character encoding.

omh-live-info