The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to capture and store user intent for retrieval by future AI agents, which creates a potential surface for indirect prompt injection.
Ingestion points: Untrusted user thoughts are ingested via explicit commands or conversation monitoring and written to the project-local .ustht/ directory using scripts/write_raw.py.
Boundary markers: Recorded thoughts are stored as markdown list items under date-based headers (e.g., ## yyyy-mm-dd). There are no specific isolation delimiters or warnings within the stored data to instruct future agents to ignore embedded instructions.
Capability inventory: The skill possesses read, write, and bash capabilities. The provided Python scripts use these tools to manage the project memory base, perform maintenance (sortin/resort), and display stored content.
Sanitization: The skill implements strong security best practices, including robust path traversal protection in scripts/common.py using Path.resolve() and relative_to() checks. Additionally, scripts/write_raw.py strips newlines from captured thoughts to prevent structural injection into the raw markdown files.

user-thoughts