The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run() in scripts/normalize.py to execute document converters and custom parsers. The commands are constructed using static configuration from parsers/registry.json or well-known binary paths (e.g., markitdown). User input {input} is correctly handled as a path argument.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloading documents from the internet via the /notebook:ingest command. This process includes explicit security controls: the agent must confirm the specific domain with the user, temporarily whitelists only that domain in .claude/settings.local.json, and immediately re-seals the environment after the download is complete. This follows the principle of least privilege for network access.
[SAFE]: The 'Homoglyph' warning flagged in references/parsers.md is a false positive related to Cyrillic characters used for instructional purposes in a Russian-language document; it does not represent a typosquatting attack on a domain.
[SAFE]: The permission rules written to .claude/settings.local.json in scripts/isolation.py serve to restrict the agent's capabilities (denying WebFetch, WebSearch, and shell-based network tools like curl and wget) to ensure session privacy, which is a defensive security feature rather than a threat.

notebook