The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches untrusted web content through search snippets and full-page scrapes. This content could potentially contain instructions intended to influence the agent's behavior.
Ingestion points: Data enters the agent's context through the web_search and enrich_content tools.
Boundary markers: Output is formatted using the TOON protocol to distinguish tool results, though explicit instructions to ignore embedded commands are not present in the skill text.
Capability inventory: The skill possesses network communication capabilities and file system access for configuration and usage tracking.
Sanitization: The skill relies on formatting and Jina Reader for content processing; no specific prompt-level sanitization for adversarial instructions is documented.
[DATA_EXPOSURE]: The skill manages optional API keys for various search providers and tracks usage statistics in a local configuration directory (~/.config/seti/usage.json). These are standard operational procedures for a local-first search utility.
[EXTERNAL_DOWNLOADS]: The setup process (seti setup) facilitates the installation of the SearXNG backend using Docker or the uv Python tool manager. These operations involve downloading legitimate software images and packages from official registries.

seti