The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the external cass binary and several included shell and Python scripts.
Evidence found in SKILL.md and the scripts/ directory (e.g., scripts/recover.sh, scripts/prompt_miner.py).
[COMMAND_EXECUTION]: Provides functionality for multi-machine search using ssh and rsync to execute queries and sync data from remote hosts.
Detailed in references/REMOTE_SOURCES.md and implemented in scripts/multi_machine_search.sh.
[EXTERNAL_DOWNLOADS]: Documents a dependency on an external cass binary and supports downloading model bundles from HuggingFace via the cass models install command.
[PROMPT_INJECTION]: Presents an indirect prompt injection surface as the agent is instructed to ingest and process untrusted historical conversation data from session logs.
Ingestion points: Reads .jsonl session files from Claude Code, Codex, and Gemini CLI directories (e.g., ~/.claude/projects/) as described in SKILL.md and references/SESSION_FORMATS.md.
Boundary markers: Instructions emphasize structured extraction using jq and line-number filtering to isolate user prompts.
Capability inventory: Extensive file system read access for session logs, shell execution for tool commands, and network connectivity via ssh.
Sanitization: Uses jq for parsing structured JSONL data to reduce risks associated with raw text ingestion.

cass