The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill directs the user to add a marketplace source and install a plugin from an external repository (jarrodwatts/claude-hud) that has no established trust relationship with the environment.
[REMOTE_CODE_EXECUTION]: The installation and setup process involves executing third-party code within the Claude Code environment via commands like /plugin install and /claude-hud:setup.
[METADATA_POISONING]: There is a significant discrepancy between the declared author in the skill manifest, the claimed creator in the body text (ara.so), and the actual installation source (jarrodwatts). This deceptive layering of identity can mislead users regarding the origin and safety of the software.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The plugin continuously monitors the transcript JSONL file, which stores real-time activity from all tools and subagents used in the session.
Boundary markers: The skill does not define any boundary markers or instructions to ignore malicious data embedded in tool outputs displayed by the statusline.
Capability inventory: The plugin parses internal session logs and updates a terminal statusline every 300ms, giving it access to all data exchanged during the session.
Sanitization: There is no evidence that the plugin sanitizes the transcript data before rendering it to the terminal.
[COMMAND_EXECUTION]: The troubleshooting and requirement sections provide manual shell commands for setting up temporary directories and using system package managers (winget) to install runtime dependencies.

claude-hud-statusline