The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads and installs dependencies from public repositories during the setup process. The setup.sh script installs @openai/codex from the npm registry and the httpx, arxiv, and requests packages from PyPI. These resources are necessary for the skill's infrastructure and are sourced from trusted, well-known registries.
[COMMAND_EXECUTION]: The skill uses subprocesses to interact with system tools and external CLI applications as part of its core functionality. tools/watchdog.py executes tmux, screen, nvidia-smi, and du for server-side task monitoring. The MCP servers in mcp-servers/ execute the claude and gemini command-line interfaces to facilitate the cross-model adversarial review workflow.
[REMOTE_CODE_EXECUTION]: The Gemini and Claude review servers utilize sys.executable to spawn background worker processes via subprocess.Popen. This is a standard architectural pattern for managing asynchronous job execution within this skill's research pipeline.
[SAFE]: The skill fetches research papers from external sources like arxiv.org and semanticscholar.org, which introduces a surface for indirect prompt injection. 1. Ingestion points: tools/arxiv_fetch.py and tools/semantic_scholar_fetch.py. 2. Boundary markers: Not explicitly implemented in the fetching scripts. 3. Capability inventory: The skill has command execution capabilities via subprocesses in its monitoring and review server components. 4. Sanitization: The fetched content is not specifically sanitized for adversarial prompts, but the ingestion is a primary intended function and operates within standard agent safety boundaries.

aris-infra