The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires the installation of a global NPM package @jackwener/opencli and a custom Chrome extension to function. It also supports dynamic installation of additional CLI tools like gh, vercel, and docker via the opencli install command.
[COMMAND_EXECUTION]: The skill makes extensive use of the opencli command to perform actions on websites and desktop apps. It also executes npm install, npx tsx, and npm update commands.
[DATA_EXFILTRATION]: The skill is designed to access and read private data from sessions where the user is already logged in. Commands like opencli ones token-info, opencli twitter bookmarks, opencli facebook friends, and opencli quark ls access session tokens, private messages, and file structures. This creates a high surface area for potential data exposure to the agent's context.
[REMOTE_CODE_EXECUTION]: The skill implements a 'Self-Repair' feature (via opencli-autofix) that instructs the agent to modify the source code of adapters at RepairContext.adapter.sourcePath and re-execute them. This facilitates the generation and execution of dynamic code at runtime.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests untrusted content from platforms like Reddit, Twitter, and general websites (opencli web read).
Ingestion points: Data enters the context via website adapters in SKILL.md (e.g., bilibili, twitter, reddit, web read).
Boundary markers: None identified in the provided instructions to separate untrusted web content from agent instructions.
Capability inventory: The skill has destructive and communicative capabilities including twitter delete, quark rm, douyin publish, and boss send.
Sanitization: No sanitization or validation of the ingested external content is mentioned.

qiaomu-opencli-usage