The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs local system maintenance using standard shell commands such as du, df, find, and rm.
All destructive operations are protected by a mandatory manual approval phase (Phase 3), ensuring no data is removed without explicit user consent.
The instructions advocate for safer practices, such as using rm without the -f flag and prioritizing official cleanup tools like docker image prune or pnpm store prune over manual deletion.
[DATA_EXFILTRATION]: The skill investigates disk usage patterns to report capacity changes to the user.
No network-related tools (such as curl or wget) or external communication patterns were detected, confirming the analysis remains entirely local.
[PROMPT_INJECTION]: The skill ingests file system metadata, such as directory names and paths, which are then processed by the agent.
The risk of indirect prompt injection via malicious filenames is mitigated by the skill's design, which requires findings to be presented in a structured table for human verification before any action is taken.

cleanup-storage