The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the local wx command-line tool. It explicitly notes that these commands require dangerouslyDisableSandbox: true because they must access directories outside the default agent sandbox, such as ~/.wx-cli and the WeChat container directory on macOS.
[DATA_EXFILTRATION]: The skill accesses sensitive chat data from the local WeChat installation to generate summaries. This access is inherent to the skill's purpose. The instructions include specific privacy guardrails to prevent the inclusion of sensitive personal information (phone numbers, real names, medical info) in the generated profiles and digests.
[PROMPT_INJECTION]: The skill includes dedicated instructions in references/group-memory.md to identify and ignore malicious instructions embedded in chat messages (Indirect Prompt Injection). It specifically filters for attempts to manipulate the bot's behavior through chat material.
[EXTERNAL_DOWNLOADS]: The skill references a third-party tool (wx-cli) and provides installation instructions via npm. It explicitly forbids the agent from automatically installing the tool or executing sudo commands on the user's behalf.

baoyu-wechat-summary