The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill relies on the agent's ability to interpret the file system and task history to distinguish between 'results' and 'temporary files.' This creates an attack surface where an adversary could craft filenames or directory structures that might influence the agent to delete unintended files or skip the deletion of malicious ones.
Ingestion points: File system paths, directory metadata, and task session logs are read to identify candidates for deletion (SKILL.md).
Boundary markers: The instructions explicitly warn the agent not to judge files based solely on extensions, names, or 'git untracked' status, which acts as a partial mitigation.
Capability inventory: The skill directs the agent to perform file deletion and content comparison (hashing/content checks).
Sanitization: The skill mandates that only 'precise paths' from a verified list be deleted and prohibits the use of broad wildcards or recursive deletion of unverified directories.
[COMMAND_EXECUTION]: The skill's primary purpose is to orchestrate file deletion operations. Although the instructions include constraints such as avoiding broad wildcards and verifying ownership, any automated destructive operation carries an inherent risk of data loss if the agent incorrectly attributes file ownership or misinterprets the task boundary.

finish-cleanly