The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill performs a global installation of the qmd tool directly from an unverified GitHub repository (github:tobi/qmd). Installing and running code from personal repositories allows for arbitrary code execution with the user's system privileges.
[COMMAND_EXECUTION]: The skill executes local JavaScript files (scan-vault.cjs and check-availability.cjs) assumed to exist within the user's vault. These scripts are not part of the skill's distributed code, making their behavior unverifiable. Additionally, the skill modifies sensitive configuration files like ~/.claude.json to register MCP servers.
[EXTERNAL_DOWNLOADS]: The skill downloads and executes the Bun runtime installer from a remote source (https://bun.sh/install) via a pipe to bash. While the source is a well-known service, this method is a high-risk operation that bypasses traditional package verification. It also downloads approximately 2GB of AI models from HuggingFace.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its vault scanning feature.
Ingestion points: Vault files processed by the scan-vault.cjs script.
Boundary markers: Absent; the agent treats the scanner's JSON output as trusted data for configuration.
Capability inventory: Execution of shell commands for tool installation (bun, brew) and index management (qmd).
Sanitization: Absent; metadata and paths extracted from the vault are interpolated directly into command arguments and configuration files.

enable-semantic-search