The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script tools/feishu_mcp_client.py uses npx -y feishu-mcp --stdio to dynamically download and execute the feishu-mcp package from the NPM registry at runtime. Running unpinned packages from public registries introduces a supply chain risk.
[COMMAND_EXECUTION]: The main SKILL.md uses the Bash tool to execute a suite of local Python scripts for data parsing, collection, and version management. Additionally, tools/feishu_mcp_client.py uses subprocess.run to call external Node.js utilities.
[EXTERNAL_DOWNLOADS]: The skill is designed to fetch large volumes of data from external enterprise APIs (Feishu/DingTalk) and via browser automation using Playwright. While this is the intended functionality, it involves handling sensitive corporate communications and documents.
[CREDENTIALS_UNSAFE]: The skill instructs users to provide enterprise App ID and App Secret for Feishu and DingTalk. These secrets are stored locally in the user's home directory (~/.colleague-skill/feishu_config.json and ~/.colleague-skill/dingtalk_config.json). While this is standard for local tools, the handling of high-privilege enterprise tokens requires careful user oversight.
[PROMPT_INJECTION]: The skill processes untrusted third-party data (chat logs, emails, documents) through its analysis prompts (prompts/work_analyzer.md, prompts/persona_analyzer.md). This represents an indirect prompt injection surface where malicious instructions hidden in a colleague's messages could attempt to influence the agent's behavior during the 'distillation' process.
Ingestion points: Reads files via Read tool, parses EML/MBOX via email_parser.py, and collects API/Browser data via feishu_auto_collector.py and dingtalk_auto_collector.py.
Boundary markers: Missing; the skill directly interpolates raw message content into prompts for analysis.
Capability inventory: Uses Write and Edit to create new skill files and Bash to execute local scripts and NPM packages.
Sanitization: No explicit sanitization or filtering of the ingested chat/email content is performed before analysis.

create-colleague