The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes ssh to execute search queries on remote machines in parallel via the scripts/multi_machine_search.sh script. This is an intended feature for multi-agent fleet management and employs secure practices by passing queries through stdin to prevent shell injection.
[EXTERNAL_DOWNLOADS]: The cass models install functionality downloads semantic model bundles from HuggingFace. As HuggingFace is a trusted organization for AI resources, this operation is considered safe and standard for the tool's semantic search capabilities.
[INDIRECT_PROMPT_INJECTION]: The skill processes past AI agent session logs, which represents the ingestion of untrusted data. The skill mitigates this risk through extensive instructional guardrails that warn against auto-adopting retrieved instructions and mandate human verification of cited outcomes.
Ingestion points: Agent session logs (.jsonl files) are read from project and user directories by cass and scripts/prompt_miner.py.
Boundary markers: The skill documentation provides explicit warnings to treat history as evidence rather than doctrine and requires citation of source paths and lines.
Capability inventory: Capabilities include file reading, network communication (SSH), and session resumption (cass resume --exec).
Sanitization: prompt_miner.py normalizes whitespace, and the cass tool uses structured JSON for all machine-readable output to maintain clear data boundaries.
[DATA_EXFILTRATION]: The skill transmits search queries to remote hosts via SSH to facilitate cross-machine search. This data transfer is limited to the user's search terms and is a primary functional requirement of the fleet-wide search feature.

cass