The Agent Skills Directory

[COMMAND_EXECUTION]: The scaffolded CLI tool bin/kb includes functionality to modify the user's crontab for scheduled synchronization tasks via the autosync command. Additionally, the wire.sh script modifies global AI agent configuration files located in the user's home directory (e.g., ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md) to inject the knowledge base protocol into every session.\n- [DATA_EXFILTRATION]: The sync.sh script performs automated git pull and push operations to a user-defined remote repository. While intended for synchronization, this mechanism could be used to transmit collected knowledge base data to an external server.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It is designed to ingest untrusted documents from the raw/ directory and follow metadata instructions to fetch external content from sources: URLs.\n
Ingestion points: templates/AGENTS.md (read protocol), templates/bin/kb (ingest command)\n
Boundary markers: Instructions in AGENTS.md advise agents to treat notes as background context, but no technical delimiters or safety filters are implemented in the scripts.\n
Capability inventory: The skill has access to git (sync.sh), file writing (scaffold.sh, render_okf_viewer.py), and cron modification (bin/kb).\n
Sanitization: No sanitization is performed on ingested markdown content, and render_okf_viewer.py explicitly skips sanitization for markdown bodies.\n- [EXTERNAL_DOWNLOADS]: The AGENTS.md protocol instructs agents to fetch live data from external URLs provided in note metadata, particularly llms.txt endpoints, to keep the knowledge base updated.\n- [REMOTE_CODE_EXECUTION]: The render_okf_viewer.py script generates a self-contained viz.html graph viewer that loads third-party JavaScript libraries (Cytoscape, Marked) from the jsDelivr CDN and renders markdown bodies without sanitization, creating a potential cross-site scripting (XSS) vector for locally viewed files.

init