The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/inventory_repos.py executes git commands using subprocess.run. It uses subcommands such as remote, log, rev-list, and status to gather metadata about project folders. While these are intended as read-only operations, they involve executing an external binary in the context of the user's workspace.- [INDIRECT_PROMPT_INJECTION]: The skill is designed to read and summarize content from local project files (READMEs, manifests, and source code) to generate inventory descriptions. This creates an attack surface where malicious text inside those files could attempt to influence the agent's summarization or subsequent behavior.
Ingestion points: The read_text function in scripts/inventory_repos.py reads data from requirements.txt, pyproject.toml, package.json, manifest.json, main.py, and README* files across all subdirectories in the target path.
Boundary markers: No specific boundary markers or "ignore instructions" delimiters are implemented when processing the text extracted from these project files.
Capability inventory: The script has the capability to traverse the local filesystem, read file contents, execute git via subprocess.run, and print data to standard output.
Sanitization: The script implements basic sanitization for CSV formula injection by prepending a single quote to cells starting with unsafe characters (=+-@), but does not have specific filters for prompt injection content within the file summaries.

workspace-inventory