The Agent Skills Directory

[PROMPT_INJECTION]: The skill's description and mode instructions explicitly direct the agent to "bypass a guard" or "comment out a guard" when triggered by specific user keywords. Although these instructions are intended for local testing and debugging, the terminology mirrors patterns used to circumvent security or logic controls.- [COMMAND_EXECUTION]: The skill relies on executing shell commands to manage its state. It utilizes grep to locate temporary markers and employs git checkout or git restore to revert file changes on the local system.- [INDIRECT_PROMPT_INJECTION]: The skill processes content from the local codebase which serves as an untrusted input source, creating a surface for indirect prompt injection.
Ingestion points: Local project files are searched and read during the "Make a temporary change" and "Undo" modes (SKILL.md).
Boundary markers: No delimiters or instructions are provided to the agent to distinguish between code logic and potentially malicious instructions embedded in comments or strings within the codebase.
Capability inventory: The agent can read files, write to the filesystem, and execute git commands via subprocesses.
Sanitization: The skill lacks mechanisms to sanitize or validate the content of files before the agent acts upon them to make modifications.

temp