The Agent Skills Directory

[COMMAND_EXECUTION]: The skill architecture centers on the dws CLI, which is invoked by numerous Python scripts (e.g., scripts/import_records.py, scripts/bulk_add_fields.py) using subprocess.run with structured argument lists. This design avoids shell interpolation risks while providing access to DingTalk APIs.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources such as chat messages, emails, and document content. The instructions in SKILL.md mitigate this risk by requiring the agent to use the --format json flag for all commands and strictly extract identifiers from command outputs rather than guessing them.
[DATA_EXFILTRATION]: While scripts/upload_attachment.py performs network requests using urllib.request to external URLs, these are pre-signed upload endpoints (OSS) generated by the dws CLI specifically for DingTalk attachment storage. This is legitimate product behavior for handling file uploads.
[SAFE]: The helper scripts include a resolve_safe_path utility that enforces file operations to stay within the directory defined by the OPENCLAW_WORKSPACE environment variable. This effectively prevents the agent from performing path traversal attacks to access sensitive files outside the intended project directory.

dws