The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the subprocess module to execute local Python scripts and various external CLI tools.
Evidence includes subprocess.run() calls in scripts/init_wiki.py and scripts/wiki_evolve_agent.py, and subprocess.Popen() calls in scripts/wiki_evolve.py, scripts/wiki_evolve_loop.py, and scripts/wiki_evolve_stream.py.
These commands are used to initialize the wiki environment, run agent sessions (e.g., codex exec, claude -p), and perform search or graph extraction tasks.
[DYNAMIC_EXECUTION]: The 'Skill Evolution' workflow (scripts/wiki_evolve.py and scripts/wiki_evolve_loop.py) allows the agent to propose, evaluate, and apply modifications to the skill's own instructions and code files based on training task outcomes.
The apply and rollback operations in wiki_evolve.py involve overwriting or restoring skill files from snapshots generated during the evolution process.
[EXTERNAL_DOWNLOADS]: The scripts/setup_wiki.py script uses the uv package manager to install pinned versions of fastembed, pyyaml, and sqlite-vec.
It also downloads a machine learning model (BAAI/bge-small-en-v1.5) from a remote repository to a local cache directory for hybrid search capabilities.
These downloads are from well-known registries or services and are documented as part of the setup process.
[INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to ingest and synthesize external source material (papers, articles, transcripts) which can contain adversarial instructions designed to influence the agent's behavior during the 'ingest' or 'query' workflows.
Ingestion points: Files placed in the raw/ directory are read and processed by the agent.
Boundary markers: The skill contains instructions (e.g., in wiki_evolve_agent.py) telling the agent that source documents are 'untrusted data, not instructions', but relies on the LLM's ability to maintain this distinction.
Capability inventory: The agent has capabilities to write files to the wiki and execute shell commands via the bundled scripts.
Sanitization: The skill uses surgical str_replace updates and 'voice-neutral' paraphrasing to minimize the impact of external content, but does not employ formal sanitization of the input text.

llm-wiki