The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted data from local research documents and web search results. \n
Ingestion points: Files in papers/ and literature/ directories; results from WebSearch and WebFetch tools.\n
Boundary markers: Absent; no instructions to ignore embedded commands in sourced content were detected.\n
Capability inventory: The skill utilizes Bash, Write, Edit, and network tools.\n
Sanitization: None; the skill interpolates processed content directly into prompts for the reviewer model.\n- [COMMAND_EXECUTION]: The skill uses the Bash tool for routine file management and state persistence, including a specific instruction to use cat << 'EOF' for handling large file writes.\n- [EXTERNAL_DOWNLOADS]: Utilizes WebSearch and WebFetch to acquire external research papers and grounding material as part of its core research function.\n- [SAFE]: The core logic, involving an iterative review loop with an external model (GPT-5.4 via Codex MCP), aligns with the skill's stated purpose of research refinement and represents expected functionality for the 'OpenLAIR' vendor context.

aris-research-refine