The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute various shell commands and Python scripts to manage the research process. Specifically, it runs scripts/validate_ledger.py and scripts/eval_report.py to enforce data verification and quality metrics. It also utilizes CLI tools like curl, gh (GitHub CLI), and yt-dlp for data retrieval and analysis. Potential risks are mitigated by input sanitization in the orchestrator.py script before arguments are passed to shell commands.\n- [EXTERNAL_DOWNLOADS]: The documentation in tool_strategy.md suggests that the agent should install external Python packages such as yt-dlp, curl_cffi, and feedparser using pip to handle specific media and scraping tasks. While these are common tools for the specified purpose, they represent unpinned external dependencies.\n- [INDIRECT_PROMPT_INJECTION]: As a tool designed for deep web research, the skill ingests a large volume of untrusted data from the public internet. This creates a surface for indirect prompt injection attacks where malicious content on a webpage could attempt to influence the agent's behavior.\n
Ingestion points: Untrusted data enters the agent context through WebFetch, WebSearch, and curl operations targeting arbitrary URLs derived from search results.\n
Boundary markers: The skill emphasizes data isolation via a multi-phase ledger system and verification gates, but does not provide explicit technical delimiters (e.g., unique XML tags or nonces) to isolate untrusted content from the agent's internal instructions during the synthesis phase.\n
Capability inventory: The agent possesses capabilities to execute shell commands, run local Python scripts, and write to the filesystem, which increases the potential impact of a successful injection.\n
Sanitization: The analysis did not find explicit sanitization or escaping mechanisms for the text extracted from external websites before it is processed by the LLM in the synthesis phase.

insane-research-main