office secretary

A digital administrative assistant for Microsoft 365 (Outlook & OneDrive).

What to consider before installing: - The skill appears to do what it says (read/triage mail, find calendar slots, list old OneDrive files, post to Teams) and the Python code calls only Microsoft Graph endpoints. - However the registry metadata does not declare the two required environment variables (SECRETARY_CLIENT_ID, SECRETARY_TENANT_ID) shown in SKILL.md and used by the code. This is likely an oversight but could cause the platform not to surface the credential requirements to you—treat it as suspicious until corrected. - The Azure app will need delegated permissions that allow reading/modifying email and files and sending Teams messages. These are powerful rights (Mail.ReadWrite and Files.ReadWrite can modify or delete content). Only grant them for an account/tenant you trust and consider using a dedicated service account with limited data access. - The tool uses interactive authentication and stores a local token_cache.bin. Ensure that file is stored securely (it is gitignored here) and that file-system permissions meet your policy. Review and rotate tokens if you later uninstall. - If you decide to run: inspect the code yourself, register the Azure app in a restricted test tenant, grant only the minimum scopes you accept, and run in an isolated or least-privilege account first. - Fixes that would increase trust: update the registry manifest to explicitly declare required env vars and a clear primary credential, and add NOTES in SKILL.md about token_cache location and exact consent prompts.

Type: OpenClaw Skill Name: office-secretary Version: 3.1.0 The skill is classified as suspicious due to the combination of broad Microsoft Graph permissions (`Mail.ReadWrite`, `Calendars.ReadWrite`, `Files.ReadWrite`, `ChatMessage.Send`) and the direct use of command-line arguments (`sys.argv`) for parameters like `team_id`, `channel_id`, and `msg` in `secretary_engine.py` without explicit input validation or sanitization. While the code's intent appears benign and aligned with its stated purpose, these factors introduce potential vulnerabilities (e.g., path injection or malformed requests) if the OpenClaw agent or user provides untrusted input, elevating the risk profile beyond a truly benign skill.