The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script and standard git commands (e.g., git ls-files) to inventory project documents and resolve root paths. These commands use hardcoded arguments and are restricted to the local filesystem without network interaction.
[INDIRECT_PROMPT_INJECTION]: The skill processes project-level Markdown files that may contain instructions, creating a surface for indirect prompt injection where malicious content in a document could attempt to influence the agent's audit behavior. 1. Ingestion points: Reads .md, .mdx, and .mdc files across the repository, including agent-generated plans and output folders. 2. Boundary markers: The skill instructions and referenced cleanup-rubric.md provide strict evaluation criteria and required result formats. 3. Capability inventory: The skill can trash files, apply patches, and execute a local inventory script. 4. Sanitization: No explicit content sanitization is performed; safety relies on the manual invocation guard and specific task instructions.

audit-memories