The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/upstream_check.py script executes system commands like curl and pip using subprocess.run. The implementation uses list-based arguments and avoids shell=True, which is a secure practice to prevent shell injection from untrusted input.
[EXTERNAL_DOWNLOADS]: The skill fetches skill definitions and version metadata from GitHub repositories using curl. These downloads are used to compare local files with remote upstream sources to detect updates.
[PERSISTENCE]: The skill provides instructions and a dedicated script for setting up a recurring cron job (e.g., monthly) to automate drift detection across the skill library. This establishes a persistence mechanism on the host system to ensure regular execution.
[INDIRECT_PROMPT_INJECTION]: The skill includes a 'Parallel diff dispatch' workflow that fetches markdown content from external, user-defined repositories and processes it using subagents for diffing and classification. This introduces a surface for indirect prompt injection where malicious instructions embedded in an upstream repository could influence the agent's behavior during the synchronization process.
Ingestion points: The UPSTREAM_MANIFEST.md file (which specifies sources) and the remote SKILL.md files fetched from those sources (e.g., via raw.githubusercontent.com).
Boundary markers: No explicit delimiters or instructions are provided to the subagents to ignore potentially malicious or instructional content within the fetched markdown.
Capability inventory: The agent has capabilities for writing files, running the maintenance script, and executing system commands defined in the manifest or script configuration.
Sanitization: While the script uses regex for metadata extraction, the agent-driven workflow processes the raw markdown content without formal sanitization or filtering of embedded instructions.

skill-maintainer