The Agent Skills Directory

Command Execution: The skill utilizes a local Python script (extract.py) to facilitate the extraction of log bundles. This script is part of the skill's internal logic and is implemented with significant security guardrails, including validation against path traversal (Zip Slip), resource exhaustion (Zip Bombs), and the use of restricted file types like symlinks.
Indirect Prompt Injection: The primary function of the skill involves processing external data from log files such as events.jsonl and agenthost.log. Because these logs may contain content from previous user interactions or system events, they represent a potential surface for indirect prompt injection. The skill mitigates this by instructing the agent to keep analysis local and providing specific context on sensitive data handling.
Ingestion points: Files extracted from zip archives or folders, specifically events.jsonl, usage.jsonl, ahp/*.jsonl, and various .log files.
Boundary markers: The SKILL.md file provides operational boundaries, instructing the agent to avoid quoting secrets and to maintain local analysis, though it does not define explicit data delimiters for the LLM.
Capability inventory: The skill uses shell execution for the extraction script (python3) and performs file read operations on the resulting logs.
Sanitization: The extract.py script performs extensive sanitization of the archive structure (filenames, paths, and metadata) to prevent system-level exploitation during extraction.

agent-host-logs