Loading ToolBench
Google Drive tools, at the file boundary: finding files, where they live, who can see them, folders, transfers, sharing and history.
Strong foundation with 32 well-named tools following verb_noun convention (get_, list_, create_, update_, delete_, search_). All tools have descriptions (avg ~80 chars, within baseline 34-392). Input schemas present with typed parameters and descriptions. However, output schemas are not documented in source code, only inferred from test files. Parameter descriptions are minimal (often single phrase) and lack format/constraint details. No enum constraints visible for multi-value fields like 'role' in share_file. Error handling and recovery guidance absent from tool definitions. Security-sensitive tools (delete_file, share_file, manage_approval) lack explicit permission/confirmation patterns.
Add a comment to a file.
Copy a file.
Create a new file.
Create a new folder.
Create a shortcut to a file.
Permanently delete a file.
Download a file to the local filesystem.
Output schemas not documented. Tool definitions show input schemas but no return type documentation. LLMs cannot plan downstream calls or extract required fields (e.g., does get_file return folder_id for chaining to list_folder?).
Enum constraints missing for multi-value parameters. share_file 'role' parameter lacks enum (reader|commenter|writer). manage_approval 'action' parameter lacks enum (approve|decline|cancel|comment|reassign). LLMs may hallucinate invalid values.
Destructive operations lack confirmation/dry-run pattern. delete_file, trash_file, and manage_approval (decline action) are irreversible but have no confirmation step or dry-run mode. Agents can cause data loss without safeguards.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2025-06-18+ | v2 |
Get the signed-in account and the tools this server registered.
Get a file by id, URL or path. Returns the file's metadata, folder, and who can reach it.
List pending access requests for a file.
List approvals (file reviews) on a file.
List recent changes to files in a drive.
List comments on a file.
List shared drives the account can access.
List the contents of a folder by id or path.
List who can access a file and how.
List the revision history of a file.
Manage file approvals (reviews).
Manage shared drive settings.
Manage file revisions.
Move a file to a different folder.
Read the text content of a file.
Reply to a comment.
Approve or deny an access request.
Restore a file from trash.
Search for files by name or text. Returns matching files with their locations and access.
Share a file with users or groups.
Move a file to trash.
Remove access to a file.
Update the text content of a file.
Update file metadata (name, description, etc).
Upload a file from the local filesystem.
Parameter descriptions lack format/constraint details. 'limit' parameters have no min/max bounds (e.g., list_folder limit could be 1 - 1000?). 'path' parameters lack format guidance (absolute vs relative, URL encoding). LLMs cannot validate inputs.
No error recovery guidance in tool descriptions. Tools lack hints like 'If file not found, try search_files()' or 'If permission denied, check list_permissions()'. Agents cannot self-correct on failures.