The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation and implementation reference an external dependency, @tobilu/qmd. It suggests users perform setup and indexing via /mnemonic:qmd-setup, which involves installing or utilizing unverified third-party software outside of a controlled environment.
[COMMAND_EXECUTION]: Several search patterns provided (e.g., rg -i "search term", qmd search "auth") directly interpolate user-provided strings into shell commands. This creates a surface for command injection if the agent does not adequately sanitize shell metacharacters in the user's query.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests and 'synthesizes' content from memory files which may be influenced by external data or previous untrusted interactions.
Ingestion points: The skill reads .memory.md files from the ${MNEMONIC_ROOT} directory using the Read and Bash tools.
Boundary markers: None are defined; the instructions do not require the agent to wrap memory content in protective delimiters or ignore embedded instructions during the synthesis phase.
Capability inventory: The skill has broad capabilities including Bash for command execution, Write for file modifications, and Task for initiating sub-agent calls (mnemonic-search-subcall).
Sanitization: There is no evidence of sanitization, filtering, or validation of the content retrieved from memory files before it is processed by the agent or passed to synthesis tasks.

search