The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes session logs (JSONL session log on disk) which contain untrusted conversation history, creating a surface for indirect prompt injection.
Ingestion points: The skill reads session log files from the host's disk to aggregate token counts and mode transitions.
Boundary markers: The instructions specify that the report should be printed inside a fenced code block. While this provides a visual boundary, the explicit instruction to "print it verbatim" and "do exactly that" may pressure the model to output malicious payloads if they are embedded in the logs.
Capability inventory: The skill uses a JavaScript hook (src/hooks/caveman-stats.js) to read local files.
Sanitization: The provided documentation does not mention any sanitization or validation of the log content before it is passed to the agent as additionalContext.
[COMMAND_EXECUTION]: The skill relies on a hook mechanism that executes a local JavaScript file (src/hooks/caveman-stats.js) on the host system whenever the /caveman-stats command is invoked.

caveman-stats