The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess calls in scripts/init_wiki.py to manage local runtime setup and in scripts/wiki_evolve.py to interface with various agent CLIs (such as claude, gemini, and codex). These calls are used to perform the skill's primary functions, such as orchestrating multi-agent evaluations and bootstrapping the environment.
[EXTERNAL_DOWNLOADS]: The scripts/setup_wiki.py utility employs uv to install pinned dependencies including fastembed, sqlite-vec, and pyyaml. It also downloads pre-trained embedding models from Hugging Face (BAAI/bge-small-en-v1.5). These operations target reputable and well-known services.
[DATA_EXFILTRATION]: In its evolution and evaluation workflows, the skill interacts with the Anthropic API (api.anthropic.com) through scripts/wiki_evolve_api.py. This is an established and trusted service required for the skill's evaluation logic.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted external sources like PDFs and web articles. The documentation explicitly acknowledges this risk, instructing the agent to treat these sources as untrusted data rather than instructions and providing a workflow focused on summarization and surgical edits to mitigate potential injection attempts.
[DYNAMIC_EXECUTION]: The 'Skill Evolution' feature utilizes a loop to generate code improvements (proposals) based on task experience. While this allows for modifying skill files, the process is governed by human-in-the-loop review (the apply command) and execution within isolated workspaces to ensure safety and transparency.

llm-wiki