The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted source code as data (SKILL.md), which creates a surface for indirect prompt injection via malicious comments or strings. The risk is mitigated by a mandatory evidence chain: (1) Ingestion points include code files and directory scans; (2) Boundary markers are absent in instructions, but (3) Capabilities are limited to local file modification and git history checks; (4) Sanitization is replaced by a rigorous process of behavioral preservation (铁律一), mandatory test passing (铁律五), and a human-in-the-loop confirmation gate (确认闸门) before any action is taken.
[COMMAND_EXECUTION]: The instructions utilize standard development tools such as git blame and git log to analyze code history. These operations are performed within a controlled workflow and do not involve arbitrary command execution or privilege escalation.
[REMOTE_CODE_EXECUTION]: The skill mentions the use of codemods and AST transformations for large-scale changes. These are standard industry tools for automated refactoring and are subject to the skill's requirements for test verification and user approval.

code-simplify