The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute external CLI tools using flags that intentionally disable safety protocols. Specifically, it uses the --yolo flag for Cursor Agent and the --dangerously-skip-permissions flag for Claude Code, both of which bypass human-in-the-loop confirmations.
[DATA_EXFILTRATION]: The instructions require the agent to search user terminal configuration files (shell profiles like .bashrc, .zshrc, or PowerShell profiles) to resolve command aliases. This process exposes the contents of these sensitive configuration files, which often contain environment variables, API keys, or system paths.
[COMMAND_EXECUTION]: On Windows platforms, the skill mandates execution via powershell.exe -Command, a technique that can be used to bypass certain execution policies or obfuscate command intent.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface for indirect injection by aggregating outputs from external agents (Kimi, Cursor, Claude) and using that unvalidated data to revise the primary plan.
Ingestion points: Review results collected from external CLI outputs in the '修订计划' (Revise Plan) phase.
Boundary markers: None. The results are processed directly to update the plan.
Capability inventory: Shell command execution via powershell.exe and native CLI calls.
Sanitization: None detected. The agent is instructed to '逐条核实' (verify item by item) but lacks programmatic sanitization of the external content.

plan-review